Method and system for predicting mining subsidence of complex working face mine

Abstract

The invention relates to the technical field of mining subsidence prediction, and discloses a complex working face mining subsidence prediction method and system.The complex working face mining subsidence prediction method comprises the steps that a geological feature coding network is constructed, and complex geological information is converted into digital feature vectors; constructing a subsidence parameter inversion model based on deep learning; the problem of insufficient new mining area data is solved by applying a transfer learning framework; analyzing and optimizing model performance through parameter sensitivity; efficient subsidence calculation is realized based on a non-uniform block model; according to the method, the limitation that a traditional subsidence prediction method depends on empirical parameters is broken through, the prediction precision under the complex geological condition is improved, the dependence on monitoring data is reduced, the calculation efficiency of a large-scale block model is improved, the multi-working-face superposition subsidence effect can be accurately simulated, and the method is suitable for being popularized and applied. Powerful support is provided for safe mining of mines and environmental protection.

Description

Technical Field

[0001] The present invention relates to the technical field of mine subsidence prediction, and more particularly to a method and system for predicting mine subsidence in a complex working face. Background Art

[0002] With the large-scale mining of coal resources, the problem of surface subsidence has become increasingly prominent, causing serious impacts on surface buildings, water resources and the ecological environment. Accurately predicting surface subsidence caused by mining is of great significance for guiding mine safety and protecting the surface environment.

[0003] Traditional probability integral subsidence prediction models require a large number of key parameters, such as the subsidence coefficient, horizontal movement coefficient, inflection point offset, and the tangent of the main influence angle. These parameters are typically determined through empirical formulas or analogical methods. In mining areas with complex geological conditions, this parameter acquisition method is subject to significant subjectivity and uncertainty, resulting in deviations between prediction results and actual conditions. Furthermore, traditional methods struggle to effectively address subsidence prediction in complex geological conditions, such as complex stratigraphic structures, variable lithologic distributions, and dense fault distributions. Furthermore, they are unable to establish a precise mapping between geological characteristics and subsidence parameters.

[0004] Existing technologies face significant challenges when working with newly mined areas or those with limited monitoring data. Traditional methods rely heavily on extensive historical monitoring data for parameter inversion, making accurate predictions difficult in the absence of sufficient data.

[0005] In addition, multi-face mining is a common mode of modern mining production. However, traditional methods are computationally complex and lack accuracy when dealing with the prediction of superimposed subsidence of multiple working faces. It is difficult to accurately simulate the impact of mining sequence on subsidence development, and it is impossible to provide timely and accurate technical support for safe mining decisions in mines.

[0006] Therefore, it is of great theoretical significance and practical value to develop a new subsidence prediction method that can adapt to complex geological conditions, improve prediction accuracy, reduce data dependence and improve computational efficiency. Summary of the Invention

[0007] The present invention provides a method and system for predicting mining subsidence in complex working faces, which solves the technical problems in related technologies such as complex calculations and insufficient accuracy when processing superimposed subsidence predictions of multiple working faces, and difficulty in accurately simulating the impact of mining sequence on subsidence development.

[0008] The present invention provides a method for predicting mining subsidence in a complex working face mine, comprising the following steps:

[0009] Collect and standardize the original geological information of the mining area, build a multi-level geological feature coding network, convert the geological information into high-dimensional feature vectors, and integrate local, medium-range and global geological features;

[0010] Combine high-dimensional feature vectors with historical mining data to train a deep learning model, establish a mapping relationship between geological features and subsidence parameters, integrate physical constraints to ensure the rationality of the prediction results, and use a transfer learning framework to transfer the trained model knowledge to the target mining area;

[0011] The model migrated to the target mining area is fine-tuned using limited field monitoring data to achieve accurate parameter inversion. Based on the current model prediction results, an active learning strategy is used to optimize the layout of subsequent monitoring points, improve the representativeness and information content of the newly added monitoring data, and combine parameter sensitivity analysis to optimize the selection and adjustment of subsidence parameters.

[0012] For the optimized subsidence parameters and models, through the incremental learning mechanism, as new monitoring data continues to accumulate, the model is continuously updated and optimized to improve the long-term adaptability and prediction accuracy of the system, and ultimately improve the overall accuracy of subsidence prediction.

[0013] In a preferred implementation, the geological feature encoding network includes a three-dimensional convolutional neural network, a graph convolutional network and a self-attention mechanism, which are used to extract local, mid-level and global geological features respectively, and generate a comprehensive feature vector through a feature fusion mechanism.

[0014] In a preferred implementation, the deep learning model adopts a multi-task learning structure, which can simultaneously predict multiple subsidence parameters and improve the accuracy and physical rationality of the prediction through parameter sharing and physical constraints.

[0015] In a preferred implementation, the transfer learning framework includes domain difference analysis, feature alignment, hierarchical transfer, and active learning monitoring point optimization layout, which can achieve high-precision parameter inversion when monitoring data is limited.

[0016] In a preferred implementation, the active learning strategy combines uncertainty sampling with representative sampling to dynamically optimize the spatial distribution of monitoring points through comprehensive scoring, thereby improving monitoring efficiency and prediction accuracy.

[0017] In a preferred embodiment, parameter sensitivity analysis includes local sensitivity analysis, global sensitivity analysis and parameter importance assessment based on machine learning, which are used to guide the optimization of key parameters and model adjustment.

[0018] In a preferred implementation, the incremental learning mechanism includes knowledge distillation, parameter regularization, and model integration updates, which can continuously optimize model performance and prevent catastrophic forgetting as monitoring data continues to accumulate.

[0019] In a preferred implementation, a method for predicting mining subsidence in complex working faces also includes an adaptive processing mechanism for heterogeneous geological structures, which uses recursive feature decomposition and dynamic boundary condition identification algorithms to automatically adjust the calculation accuracy and solution strategy according to the geological complexity of different regions, so that the system can maintain stability in a discontinuous geological environment.

[0020] In a preferred implementation, a method for predicting mining subsidence in a complex working face mine also includes a time-space decoupling algorithm and a dynamic influence function construction. The time series and spatial influence are separated and processed through four-dimensional tensor decomposition, and combined with nonlinear time factor correction, to solve the problem of subsidence evolution prediction under different mining sequences of multiple working faces, and realize rapid evaluation of different mining sequence scenarios.

[0021] In a preferred embodiment, a complex working face mining subsidence prediction system is used to execute a complex working face mining subsidence prediction method, comprising:

[0022] Geological data processing module, used to collect and standardize geological information and construct feature coding network;

[0023] Subsidence parameter mapping module, used to train deep learning models to establish mapping relationships between geological features and subsidence parameters;

[0024] Model fine-tuning and optimization module, used to fine-tune the model and optimize the monitoring point layout using limited monitoring data;

[0025] The incremental learning update module is used to continuously update the optimization model as new data accumulates. The modules work together to achieve high-precision prediction of subsidence in complex mining areas.

[0026] The beneficial effects of the present invention are:

[0027] Improved accuracy in determining subsidence parameters: Through geological feature coding and deep learning models, a precise mapping relationship between geological features and subsidence parameters is established, which improves the accuracy of parameter determination.

[0028] Reduced data requirements: Using a transfer learning framework, the system can invert the distribution of subsidence parameters in the entire area based on data from a small number of monitoring points, reducing monitoring costs and data collection difficulty.

[0029] Get rid of excessive reliance on experience: Breaking through the excessive reliance of traditional parameter determination methods on experience and historical data, establishing the intrinsic connection between geological characteristics and subsidence parameters through deep learning, and realizing intelligent inversion and optimization of parameters.

[0030] Significantly improved computational efficiency: The processing efficiency of non-uniform block models has been improved, meeting the real-time requirements of engineering practice.

[0031] Accurate simulation of the superposition effect of multiple working faces: Through innovative time-space decoupling algorithms and dynamic influence functions, the system can accurately simulate the impact of mining sequence on subsidence development, and the average error of multi-working face superposition subsidence prediction is reduced.

[0032] Uncertainty quantification and risk assessment: The introduction of methods such as Monte Carlo simulation and Bayesian updating enables uncertainty quantification and risk assessment of subsidence prediction results, providing more comprehensive information support for decision-making.

[0033] Strong system adaptability: Through transfer learning and incremental learning mechanisms, the system can continuously adapt to new mining conditions and data, maintaining long-term prediction accuracy and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flow chart of a method for predicting mining subsidence in a complex working face of the present invention;

[0035] Figure 2 is the pre-merger block of the present invention;

[0036] Figure 3 is the merged block of the present invention;

[0037] Figure 4 It is a part of the present invention to compare a small number of blocks before and after merging;

[0038] Figure 5 This is another part of the present invention, which is a comparison of a small number of blocks before and after merging;

[0039] Figure 6 It is the prediction result of the complex working face settlement system of the present invention;

[0040] Figure 7 It is a plan view of the complex working face settlement system prediction of the present invention;

[0041] Figure 8 It is a prediction contour map of the complex working face settlement system of the present invention;

[0042] Figure 9 This is a curvature result diagram of the complex working face settlement system predicted by the present invention;

[0043] Figure 10 It is a curvature plane diagram of the complex working face settlement system predicted by the present invention;

[0044] Figure 11 It is a two-dimensional contour map of the predicted curvature of the complex working face settlement system of the present invention. DETAILED DESCRIPTION

[0045] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0046] At least one embodiment of the present invention discloses a method for predicting mining subsidence in a complex working face mine. Figure 1 As shown, the following steps are included:

[0047] Step 1: Collect and standardize the original geological information of the mining area, build a multi-level geological feature coding network, convert the geological information into high-dimensional feature vectors, and integrate local, mid-range and global geological features;

[0048] The specific steps include:

[0049] Step 1.1, geological data collection and preprocessing;

[0050] Collect the original geological data of the mining area, including geological drilling data, physical and mechanical parameter test data, stratigraphic distribution maps, fault distribution maps, etc., and carry out systematic sorting and standardized processing. Specifically including:

[0051] Perform spatial interpolation on the drilling data to generate a three-dimensional stratigraphic model in a unified coordinate system;

[0052] Convert lithologic information into numerical characteristics, such as density, elastic modulus, Poisson's ratio, internal friction angle, cohesion, tensile strength and other physical and mechanical parameters;

[0053] Perform vector processing on the fault distribution to extract the geometric features of the fault, such as spatial position, dip, and strike;

[0054] The data is normalized so that features of different dimensions are within a unified numerical range, which facilitates subsequent model training.

[0055] Step 1.2, construct a hierarchical geological feature coding network;

[0056] According to an embodiment of the present application, it is necessary to design and construct a multi-level geological feature encoding network that can automatically extract high-dimensional feature vectors representing geological characteristics from raw geological data.

[0057] The specific structure is as follows:

[0058] Low-level feature extraction subnetwork: A three-dimensional convolutional neural network is used to process spatially distributed data and extract local geological features.

[0059] For the input geological characteristics of the i-th location Extract local features through convolution operation:

[0060]

[0061] in, Representation of low-level geological features, i.e., representation of primary geological features extracted through the network; Represent the original geological input characteristics at position i and position j, respectively, including information such as lithology, physical and mechanical parameters of the position; N(i) represents the neighborhood of position i, that is, the set of all position points adjacent to position i, which is used to capture the local geological structure relationship; Represents the low-level feature extraction convolution kernel weight matrix, which is used to extract effective features from the original geological data; Represents the bias term of low-level feature extraction, which is used to adjust the baseline value of feature extraction; σ low It is the activation function used in the low-level feature extraction network, which is used to introduce nonlinear transformation to enhance the model's expressiveness; j∈N(i) It represents the weighted summation of the features of all locations j in the neighborhood of location i to achieve the aggregation of local geological features.

[0062] Middle-layer feature fusion sub-network: A graph convolutional network is used to process the correlation between geological structures and capture the mutual influence between different geological units.

[0063] For the features of node v, the graph convolution operation updates the node features by performing weighted aggregation on the features of its neighbors.

[0064] This process takes into account the connectivity between nodes and the node degree (the number of edges connected to the node) to ensure that information flows rationally in the geological structure network.

[0065] In this way, mid-level features can effectively capture the mutual influence and spatial correlation patterns between geological units.

[0066] High-level semantic feature extraction sub-network: uses the self-attention mechanism to capture the long-distance dependencies between global geological features and generate the final high-dimensional feature vector.

[0067] The self-attention mechanism updates the representation of each element by calculating the correlation between the elements in the input sequence.

[0068] Specifically, the system first generates three sets of representations: query, key, and value. It then assigns attention weights based on the similarity between the query and key. Finally, it aggregates the values ​​based on these weights to form a high-level feature representation. This enables the model to focus on relevant geological information globally and effectively model long-range dependencies.

[0069] Step 1.3, feature vector generation and optimization;

[0070] The final geological feature vector is generated by integrating low-level, mid-level, and high-level features through skip connections and an attention fusion mechanism. This process combines the features of the three levels through feature concatenation operations and then transforms them through a feature integration weight matrix to form a feature vector that comprehensively expresses the geological characteristics.

[0071] In addition, in order to optimize the expressive power of feature vectors, this application provides a feature importance adaptive adjustment algorithm to dynamically adjust the weights of features in each dimension based on feedback from downstream tasks.

[0072] The algorithm dynamically adjusts the importance of different feature dimensions by applying the Sigmoid function to the feature vector and performing element-by-element multiplication on the weights generated, thereby obtaining the optimized geological feature vector.

[0073] The implementation process of the feature importance adaptive adjustment algorithm includes the following steps:

[0074] Initialize the adjustment parameters to small random values;

[0075] The geological feature vector is input into the adjustment network to calculate the importance weight of each dimension;

[0076] According to the error between the subsidence prediction result and the actual monitoring data, the parameters are updated and adjusted through back propagation;

[0077] Repeat the steps until the feature weights converge.

[0078] Step 1.4, verify and evaluate the feature encoding quality;

[0079] The quality of geological feature encoding can be assessed through methods such as dimensionality reduction visualization, cluster analysis, and feature importance analysis:

[0080] Use dimensionality reduction techniques such as t-SNE or PCA to map high-dimensional feature vectors into two-dimensional space and visually observe the distribution of samples under different geological conditions;

[0081] The degree of clustering of samples with similar geological conditions is evaluated by calculating the Euclidean distance or cosine similarity between feature vectors;

[0082] A gradient-based feature attribution method is used to analyze the contribution of each dimensional feature to subsequent subsidence prediction and guide the optimization direction of the feature extraction network.

[0083] Step 2: Combine the high-dimensional feature vectors with historical mining data to train a deep learning model, establish a mapping relationship between geological features and subsidence parameters, integrate physical constraints to ensure the rationality of the prediction results, and use a transfer learning framework to transfer the trained model knowledge to the target mining area;

[0084] The specific steps include:

[0085] Step 2.1, historical data set construction and expansion;

[0086] This application first needs to collect and organize historical mining area data with different geological conditions and mining situations to construct a representative training dataset:

[0087] Data collection: Collect complete cases containing geological characteristics information, mining parameters and actual subsidence monitoring data from different mining areas;

[0088] Data labeling: Based on actual monitoring data, the subsidence parameters of each mining area are obtained through inversion analysis as labels for model training;

[0089] Data augmentation: Generate more training samples through physical model simulation and parameter transformation to improve the generalization ability of the model;

[0090] Data splitting: The dataset is divided into training set, validation set and test set in a ratio of 8:1:1.

[0091] Step 2.2, multi-task learning network construction;

[0092] According to the technical solution of this application, a multi-task learning network can be constructed to simultaneously predict multiple subsidence parameters, making full use of the relationship between the parameters to improve the prediction accuracy:

[0093] Shared feature extraction layer: Receives the geological feature vector generated in step 1 and further extracts features through a multi-layer perceptron.

[0094] This process maps the optimized geological feature vector to a shared feature space through two layers of nonlinear transformation, so that subsequent parameter prediction tasks can be learned based on the same feature representation, thereby making full use of the common information between parameters.

[0095] Parameter-Specific Branches: Independent prediction branches are designed for each subsidence parameter (sinking coefficient, horizontal shift coefficient, influence angle, etc.). Each branch receives the output of the shared feature layer, performs a linear transformation using specific weights and bias parameters, and applies the corresponding activation function to obtain the predicted value of the specific subsidence parameter. This design enables the model to learn specialized characteristics of different parameters.

[0096] Parameter association constraint layer: This application introduces a physical relationship constraint system between parameters to ensure that the prediction results conform to the laws of geomechanics:

[0097]

[0098] in, represents the parameter association constraint loss, which is used to quantify the degree to which the predicted parameters deviate from physical laws. q represents the horizontal movement coefficient, b represents the inflection point offset, β represents the influence angle, f(b) represents the influence of the inflection point offset on the horizontal movement coefficient, and g(q) represents the influence of the horizontal movement coefficient on the influence angle. C1 and C2 represent the ratio of q to f(b) and the ratio of β to g(q) under standard geological conditions, respectively. λ1 and λ2 represent the weight coefficients of the first and second constraints, respectively, and are used to balance the contributions of different constraints to the total loss function. This constraint loss ensures that the predicted subsidence parameters conform to the basic laws of geomechanics.

[0099] Step 2.3, innovative training strategies;

[0100] In this application, a multi-stage training and combined loss function strategy is adopted to effectively improve the prediction accuracy of the model:

[0101] Multi-stage training strategy:

[0102] Phase 1: freeze parameters of specific branches and only train the shared feature extraction layer;

[0103] The second stage: freeze the shared feature extraction layer and train each parameter-specific branch separately;

[0104] Phase 3: Unfreeze the entire network and perform end-to-end fine-tuning.

[0105] Combined loss function design:

[0106] The total loss function consists of prediction loss, constraint loss, and regularization term:

[0107]

[0108] Among them, L total Represents the total loss function, which is used to guide the model training process; represents the weighted sum of the predicted losses of all subsidence parameters, i represents different subsidence parameter indexes, such as subsidence coefficient, horizontal movement coefficient, influence angle, etc.; w i is the weight coefficient of each parameter loss, which is used to balance the importance of different parameters in the total loss; is the predicted loss of each subsidence parameter, which is used to measure the difference between the predicted value and the true value; L is a constraint loss term used to penalize prediction results that violate physical laws and ensure that the prediction parameters satisfy the geomechanical relationship; reg is the model parameter regularization term, used to prevent the model from overfitting; c is the constraint weight coefficient, which controls the influence of the physical constraint term on the total loss; r is the regularization coefficient, which controls the influence of the regularization term on the total loss. By minimizing the total loss function, we can achieve accurate prediction of the subsidence parameters while ensuring that the prediction results conform to physical laws and have good generalization capabilities.

[0109] Dynamic weight adjustment:

[0110] The weight of each parameter prediction loss is dynamically adjusted based on the performance of the validation set. After each round of training, the loss weight in the next round of training is dynamically adjusted based on the prediction error of each parameter on the validation set.

[0111] Parameters with larger prediction errors will receive larger loss weights to concentrate the model's learning ability on improving the predictions of poorly performing parameters.

[0112] This dynamic adjustment method ensures that the model can balance the prediction accuracy of each parameter, avoiding the situation where some parameters are predicted too accurately while other parameters are predicted poorly.

[0113] Step 2.4, model evaluation and performance optimization;

[0114] This application uses multiple methods to evaluate model performance and optimize it:

[0115] Evaluation Metrics:

[0116] Mean absolute error and root mean square error: evaluate the accuracy of parameter predictions;

[0117] Coefficient of determination R 2 : Evaluate the model's ability to explain the relationship between geological characteristics and subsidence parameters;

[0118] Subsidence prediction error: Substitute the prediction parameters into the probability integral method to calculate the error compared with the actual subsidence monitoring data.

[0119] Model pruning and quantization:

[0120] Prune low-importance parameters based on sensitivity analysis to reduce model complexity;

[0121] Use mixed precision quantization to reduce model storage requirements and inference computation.

[0122] Ensemble learning improvements:

[0123] Train multiple deep learning models with different structures and fuse the prediction results through weighted averaging or voting mechanisms;

[0124] Techniques such as Dropout and Bootstrap are used to improve model robustness.

[0125] It can be seen that through the above steps, this application has established a deep learning model that can accurately predict subsidence parameters. This model not only takes into account the complex mapping relationship between geological characteristics and subsidence parameters, but also incorporates the physical constraints between parameters, thereby improving the accuracy and reliability of subsidence parameter prediction and laying the foundation for subsequent transfer learning and integrated prediction systems.

[0126] Step 3: Fine-tune the model transferred to the target mining area using limited field monitoring data to achieve accurate parameter inversion. Based on the current model prediction results, an active learning strategy is used to optimize the layout of subsequent monitoring points, improve the representativeness and information content of the newly added monitoring data, and optimize the selection and adjustment of subsidence parameters through parameter sensitivity analysis.

[0127] The specific steps include:

[0128] Step 3.1, domain difference analysis and feature alignment;

[0129] This application first analyzes the feature distribution differences between the source domain (historical mining area) and the target domain (new mining area), and uses feature alignment technology to eliminate the differences between the domains:

[0130] Analysis of field differences:

[0131] Calculate the marginal distribution difference of geological feature vectors between the source and target areas, and use statistics such as maximum mean difference or Wasserstein distance to measure the degree of difference between the two areas;

[0132] Feature visualization techniques (such as t-SNE) are used to intuitively display the distribution differences of features in the two fields.

[0133] Feature alignment method:

[0134] A domain-invariant feature extractor is applied to adjust the geological feature encoding network so that the extracted features are similarly distributed in the source and target domains.

[0135] The domain-invariant feature extractor enables the model to learn domain-shared feature representations by minimizing the distance between the feature distributions of the source and target domains.

[0136] This process calculates the expected distance between the features of source domain samples and target domain samples and minimizes this distance during training.

[0137] In addition, this application introduces an adversarial training model, which makes it impossible to distinguish which domain the extracted features come from through adversarial training of the domain discriminator and the feature extractor.

[0138] In this framework, the feature extractor attempts to generate features that make it difficult for the domain discriminator to distinguish which domain they come from, while the domain discriminator strives to accurately determine which domain the features come from.

[0139] Through this adversarial process, the feature extractor gradually learns to extract domain-invariant feature representations.

[0140] Traditional transfer learning methods usually fine-tune model parameters directly on new domain data, ignoring the distribution differences between the source and target domains, resulting in poor transfer effects.

[0141] Compared with traditional methods, the domain-invariant feature extractor proposed in this application has the following advantages:

[0142] Adaptive distribution alignment: By minimizing the distance between the feature distributions of the source and target domains, distribution alignment in the feature space is achieved, overcoming the domain shift problem.

[0143] Adversarial learning enhancement: Introducing adversarial training between the domain discriminator and the feature extractor, forcing the feature extractor to generate domain-independent feature representations, further improving the domain invariance of the features.

[0144] Gradual migration: By gradually increasing the weight of the target domain data, smooth migration from the source domain to the target domain is achieved, avoiding a sudden drop in model performance.

[0145] The specific implementation of this domain-invariant feature extractor includes:

[0146] First, build a feature extractor and a domain discriminator;

[0147] The feature extractor maps the data from the source and target domains into a common feature space;

[0148] The domain discriminator tries to distinguish which domain the feature comes from;

[0149] Through adversarial training, the feature extractor learns to generate features that are difficult for the discriminator to distinguish;

[0150] At the same time, by minimizing the distribution distance loss function, the similarity of the feature distributions of the two domains is directly constrained.

[0151] Step 3.2, hierarchical transfer learning model construction;

[0152] According to one embodiment of the present application, a hierarchical transfer learning model is constructed, and different transfer strategies are adopted according to different characteristics of the parameters:

[0153] Model layering:

[0154] Bottom-layer feature extraction network: mainly extracts general information of geological features and maintains relative stability;

[0155] Middle layer network: responsible for extracting key features shared by the domain and requires slight adjustments;

[0156] Top-level network: The layer that directly maps to subsidence parameters and needs to be adjusted specifically for the characteristics of the new mining area.

[0157] Layered migration strategy:

[0158] For the underlying network: apply feature freezing or slight fine-tuning to retain the basic feature extraction capabilities learned from the source domain;

[0159] For the middle layer network: adopt feature adaptation algorithm and introduce adaptive batch normalization technology to adjust the feature distribution to adapt to the target domain;

[0160] For the top-level network: perform complete retraining or parameter reinitialization and optimize it specifically for specific tasks in the target domain.

[0161] The advantage of this hierarchical transfer learning strategy is that it recognizes that the features learned at different layers of the neural network have different degrees of generality and specificity.

[0162] The underlying network usually learns basic features such as texture and edges, which have high commonality across different fields;

[0163] The middle layer network captures more complex feature combinations and needs to be moderately adjusted to adapt to new domains;

[0164] The top-level network is directly related to the task and needs to be significantly adjusted for the new domain.

[0165] Step 3.3, active learning of monitoring point optimization layout;

[0166] In view of the limited monitoring data, this application constructs an active learning strategy to optimize the layout of monitoring points and maximize information acquisition:

[0167] Uncertainty sampling:

[0168] The uncertainty of the prediction model is used to evaluate the information value of each potential monitoring point, and the location with the highest uncertainty is selected for monitoring;

[0169] Uncertainty calculation methods are primarily based on the entropy of the predicted probability distribution or the complement of the maximum predicted probability. The degree of uncertainty in the prediction at a location is quantified by calculating the entropy of the predicted probability distribution of the parameter at that location (i.e., the degree of uncertainty in the predicted distribution) or using the complement of the maximum predicted probability (i.e., 1 minus the maximum predicted probability). Locations with higher uncertainty contain greater potential information and are more valuable for model training.

[0170] Representative sampling:

[0171] Select key locations that can represent the overall distribution of characteristics of the mining area for monitoring to improve the representativeness of monitoring data;

[0172] The representativeness score is evaluated by calculating the feature space distance between the candidate point and the nearest point in the selected monitoring point set. The farther the candidate point is from the selected point, the higher the representativeness. This ensures the diversity of the monitoring point distribution and avoids the occurrence of redundant monitoring points.

[0173] Monitoring strategy optimization:

[0174] In addition, combining uncertainty sampling and representative sampling, this application designs a hybrid sampling algorithm:

[0175] Score(x)=λ balance U(x)+(1-λ balance )·R(x);

[0176] Among them, Score(x) is the comprehensive score of the candidate monitoring point x, which determines the priority of the point to be selected as the monitoring point; U(x) is the uncertainty score, which indicates the uncertainty of the prediction result of the location. The higher the value, the more valuable the point is for model improvement; R(x) is the representativeness score, which indicates the degree of difference between the point and the selected monitoring points. The higher the value, the more new information the point can provide. balance It is a balance factor with a value range of 0 to 1. It is used to balance the weights of uncertainty sampling and representative sampling, and is dynamically adjusted according to the characteristics of the mining area.

[0177] Traditional monitoring point placement methods typically use equidistant or empirical placement, which fails to optimize key areas for subsidence prediction. In contrast, the hybrid sampling algorithm proposed in this application achieves optimal configuration of monitoring points by dynamically balancing uncertainty and representativeness.

[0178] The specific implementation of this algorithm includes:

[0179] Mining area discretization: Divide the mining area surface into grids, with each grid point serving as a potential monitoring location;

[0180] Initial monitoring point selection: Based on prior knowledge of geological characteristics, a small number of initial monitoring points are selected;

[0181] Model prediction and uncertainty calculation: Use existing monitoring data to train the model and predict the subsidence parameters and their uncertainties for all grid points;

[0182] Balance factor determination: The balance factor is dynamically adjusted based on the sparsity and distribution of the current monitoring data. Areas with sparse monitoring points tend to increase the representative weight, while areas with high gradient changes tend to increase the uncertainty weight.

[0183] Monitoring point addition: Calculate the comprehensive scores of all candidate points and select the location with the highest score to arrange new monitoring points.

[0184] Step 3.4, incremental learning and continuous updating mechanism;

[0185] This application also designs an incremental learning mechanism to continuously update and optimize the model as monitoring data continues to accumulate:

[0186] Knowledge distillation technology:

[0187] Use the original model as the teacher model and the newly trained model as the student model, retaining the general knowledge of the original model through knowledge distillation;

[0188] The distillation loss function combines the cross-entropy loss of standard supervised learning with the KL divergence between the output distributions of the teacher and student models.

[0189] Specifically, the loss function consists of two parts:

[0190] One part is the cross entropy loss between the student model’s predictions and the true labels, and the other part is the KL divergence between the output distributions of the student model and the teacher model after adjusting the temperature parameter.

[0191] By balancing these two parts of loss, the student model can not only learn label information but also obtain rich knowledge representation from the teacher model.

[0192] Parameter regularization:

[0193] This application introduces algorithms such as Elastic Weight Consolidation (EWC) to prevent the model from forgetting its original knowledge when adapting to new data;

[0194] The EWC algorithm is implemented by adding constraints on key parameters to the loss function. The algorithm first calculates the importance of the original model parameters to the original task (represented by the Fisher information matrix), and then imposes penalties on changes in important parameters when the model is adapted to new data, so that these parameters do not deviate too far from the original values.

[0195] This approach allows the model to adapt to new data while retaining existing knowledge, effectively preventing catastrophic forgetting.

[0196] Model integration updates:

[0197] Use model integration technology to perform weighted integration of the newly trained model and the original model;

[0198] As monitoring data accumulates, the weights of each model are dynamically adjusted, and the weights of new models are gradually increased.

[0199] In step 4, for the optimized subsidence parameters and models, the incremental learning mechanism is used to continuously update and optimize the model as new monitoring data continues to accumulate, thereby improving the long-term adaptability and prediction accuracy of the system and ultimately improving the overall accuracy of subsidence prediction.

[0200] The specific steps include:

[0201] Step 4.1, local sensitivity analysis method;

[0202] This application quantitatively analyzes the impact of changes in a single parameter of the subsidence prediction model on the output results:

[0203] Single parameter perturbation analysis:

[0204] While keeping other parameters unchanged, make a small perturbation to a single parameter and calculate the rate of change of the output result;

[0205] The sensitivity coefficient is calculated by multiplying the partial derivative of the output with respect to the parameter change by the ratio of the parameter to the output. This calculation method considers the relative impact of parameter changes rather than their absolute impact, allowing for fair comparison of parameters of different magnitudes. A larger sensitivity coefficient indicates a more significant impact of the parameter on the prediction result.

[0206] The parameter perturbation range is ±10% of its standard value. The sensitivity coefficient under each perturbation is calculated, and the average value is taken as the final sensitivity evaluation value.

[0207] Gradient analysis method:

[0208] Automatic differentiation technology is used to calculate the gradient of the subsidence prediction results with respect to each parameter;

[0209] The normalized gradient index is calculated by multiplying the absolute partial derivative of the subsidence prediction result with respect to the parameter by the ratio of the parameter standard deviation to the output standard deviation.

[0210] This normalization process allows parameters of different scales to be directly compared in terms of their impact. The larger the index value, the more significant the parameter impact.

[0211] Step 4.2, global sensitivity analysis technique;

[0212] In this application, the interaction between parameters and the impact of global changes on subsidence prediction results are considered:

[0213] Variance decomposition method:

[0214] Decompose the model output variance into the contribution of each parameter and its interaction term;

[0215] The first-order sensitivity index is obtained by calculating the ratio of the individual contribution of a parameter to the output variance to the total output variance.

[0216] Specifically, for a certain parameter, fix the value of the parameter, calculate the conditional expectation of the output when other parameters change randomly, then calculate the variance of this conditional expectation with respect to the fixed parameter, and finally take the ratio of this variance to the total variance of the output as the first-order sensitivity index.

[0217] The overall sensitivity index measures the degree to which a parameter and its interaction with other parameters contribute to the output.

[0218] The calculation method is to subtract the proportion of the contribution of all parameters except this parameter to the output variance from 1.

[0219] The total sensitivity index includes not only the first-order effects of a parameter but also all higher-order interaction effects involving that parameter.

[0220] Morris screening method:

[0221] This application is based on a single-factor experimental design to calculate the basic effect of each parameter;

[0222] The mean absolute base effect and standard deviation are calculated as follows:

[0223] For the parameter Xi, add a disturbance value Δ to it, calculate the ratio of the change in the model output to the disturbance size, and get the basic effect of the parameter;

[0224] Repeat this process r times, calculate the average of the absolute values ​​of all basic effects, and obtain the average absolute basic effect;

[0225] Calculate the standard deviation of the basic effect to reflect the strength of the parameter nonlinear effect or interaction effect.

[0226] Parameters with larger average absolute basic effects have a significant impact on the model output, while parameters with larger standard deviations may have strong nonlinear effects or interaction effects with other parameters.

[0227] Step 4.3, parameter importance evaluation based on machine learning;

[0228] This application uses machine learning algorithms to evaluate parameter importance and overcome the limitations of traditional methods:

[0229] Random Forest Feature Importance Analysis:

[0230] Train a random forest model and evaluate parameter importance by calculating the impact of feature permutations on model performance;

[0231] The permutation importance calculation process is:

[0232] For each decision tree, record its original prediction error, then randomly shuffle the value of a certain feature, and then calculate the prediction error after shuffling;

[0233] The difference between the two is the importance contribution of the feature to the tree;

[0234] Finally, the results of all trees are averaged to get the overall importance score of the feature.

[0235] A higher importance score indicates that the parameter contributes more to the model's predictive ability.

[0236] SHAP value analysis:

[0237] According to the technical solution of the present application, the marginal contribution of each parameter to the prediction result is calculated based on the Shapley value of game theory;

[0238] The SHAP value calculation is based on cooperative game theory and calculates the marginal contribution of the parameters to the prediction results.

[0239] For each parameter, we consider the marginal gain in prediction when adding it to different parameter subsets and take a weighted average of all possible subsets. The weights are determined based on the size of the subsets, ensuring fairness in the calculation.

[0240] The SHAP value has a distribution property. The sum of the SHAP values ​​of all parameters is equal to the difference between the model predicted value and the benchmark value. Therefore, the SHAP value can intuitively reflect the relative importance of each parameter.

[0241] Step 4.4, parameter sensitivity visualization and optimization guidance;

[0242] Based on the sensitivity analysis results, this application constructed visualization tools and parameter optimization strategies:

[0243] Multidimensional sensitivity visualization:

[0244] Heat map: intuitively display the sensitivity coefficient of each parameter;

[0245] Radar chart: compare global and local sensitivity indicators of different parameters from multiple angles;

[0246] Interaction matrix: shows the degree of interaction between parameters.

[0247] Parameter optimization strategy formulation:

[0248] Determine the optimization priority of key parameters based on the sensitivity analysis results;

[0249] Provide more refined adjustment suggestions and optimization ranges for highly sensitive parameters;

[0250] For low-sensitivity parameters, simplified processing or default value settings can be used to reduce computational complexity.

[0251] Adaptive parameter adjustment mechanism:

[0252] This application designs a sensitivity-based adaptive parameter adjustment algorithm to dynamically optimize parameter selection;

[0253] Iterative optimization process:

[0254] Initial parameter estimation; sensitivity analysis; fine-tuning of key parameters; result evaluation; parameter update; the convergence condition is that the change in the prediction results of two consecutive iterations is less than the preset threshold.

[0255] Compared with traditional parameter adjustment methods, which usually use grid search or random search to adjust all parameters equally, the adaptive parameter adjustment algorithm of this application has the following advantages:

[0256] Priority allocation based on sensitivity: Different adjustment priorities and adjustment steps are assigned to different parameters based on their sensitivity index, concentrating computing resources on key parameters;

[0257] Dynamic parameter adjustment range: Adaptively adjust the parameter change range based on the parameter sensitivity index and the current prediction error, and use finer adjustment steps for parameters with high sensitivity;

[0258] Consideration of parameter coupling relationships: Identify the interactions between parameters through global sensitivity analysis and perform joint optimization on highly coupled parameter groups.

[0259] The specific implementation of the adaptive parameter adjustment algorithm includes:

[0260] Parameter sensitivity calculation: calculate the first-order sensitivity index and total sensitivity index of each parameter;

[0261] Parameter classification and grouping: Parameters are divided into three groups of high sensitivity, medium sensitivity and low sensitivity according to the sensitivity index, and strongly coupled parameter pairs are identified;

[0262] Iterative adjustment strategy:

[0263] For highly sensitive parameters: use a fine-grained step size for a refined search, where the step size is the benchmark step size divided by the sensitivity index of the parameter;

[0264] Alignment sensitivity parameter: Use medium step size (reference step size) for search;

[0265] For low sensitivity parameters: use a coarse step size (base step size multiplied by the sensitivity index) or keep the default value;

[0266] For strongly coupled parameter pairs: perform a two-dimensional joint search, taking into account the interaction between the parameters;

[0267] Convergence judgment: When the change in prediction error between two consecutive iterations is less than a preset threshold (usually set to 0.5%), the optimization is considered to have converged.

[0268] Application examples of this implementation:

[0269] The method of this embodiment is applied to the subsidence prediction work of a large coal mining area. The mining area covers an area of ​​about 12 square kilometers and has the following complex geological characteristics:

[0270] Complex lithology: There are five main lithologies distributed in the mining area, including sandstone, shale, siltstone, carbonaceous shale and coal seams. The thickness and physical and mechanical properties of each rock layer vary greatly;

[0271] Fault development: There are 3 major faults and 12 secondary faults distributed in the area, with fault dips ranging from 35° to 78° and complex strikes;

[0272] Large burial depth variation: The burial depth of coal seams ranges from 150 meters to 480 meters, showing a clear spatial variation trend;

[0273] The mining working faces are complex: there are 8 main working faces planned, which are planned to be mined in 3 stages, and there are overlapping impacts of multiple working faces.

[0274] The main challenges facing subsidence prediction in this mining area are limited monitoring data (only 23 monitoring points), complex geological conditions and a lack of analogous mining areas, and the low prediction accuracy of traditional methods (average error exceeding 30%). Therefore, a high-precision subsidence prediction method that can adapt to complex geological conditions and limited data is urgently needed.

[0275] Example of building a geological feature coding network:

[0276] In the implementation of this mining area, the following geological data were first collected and processed:

[0277] Drilling data: Geological data of 83 drill holes in the mining area were collected, including drill hole coordinates, thickness of each rock layer, lithology description, etc.

[0278] Physical and mechanical parameters: The test results of 28 physical and mechanical parameters of 5 main rock types were sorted out, including density (1.8 to 2.7 g / cm 3 ), elastic modulus (5.2-42.8GPa), cohesion (0.8-4.5MPa), etc.;

[0279] Fault distribution: Through geological exploration and geophysical detection, the spatial distribution data of 15 faults were obtained, including three-dimensional coordinates, dip, strike and drop.

[0280] Based on these raw data, a three-layer geological feature coding network was constructed:

[0281] Low-level feature extraction: A three-dimensional convolutional neural network is applied to process spatially distributed data. A four-layer convolutional structure is designed with convolution kernel sizes of 3×3×3, 5×5×5, 3×3×3, and 3×3×3, and the number of channels is 16, 32, 64, and 128, respectively.

[0282] The input is a 50×60×30 three-dimensional grid representing the mining area, each containing a lithology code and physical parameters. After a convolution operation, a 256-dimensional feature vector representing the local geological characteristics is successfully extracted, capturing local features such as changes in rock layer thickness and lithologic interfaces.

[0283] Mid-level feature fusion: The fault system is modeled as a graph structure, where nodes represent fault intersections and key locations (86 nodes in total), and edges represent fault connection relationships.

[0284] A three-layer graph convolutional network is designed, with hidden layer dimensions of 128, 256, and 192, respectively, and LeakyReLU is used as the activation function.

[0285] This sub-network successfully captures the topological relationship of the fault network and the mutual influence between faults, especially the modeling effect of the association pattern between the main fault and the secondary fault.

[0286] High-level semantic feature extraction: A self-attention network with a 6-head attention mechanism is used, with an input dimension of 512, 6 attention heads, and a dimension of 64 per head.

[0287] By calculating the attention weights between global geological features, the sub-network effectively identifies the influence relationships between distant geological units, especially capturing the long-range dependence of the impact of distant faults on local subsidence.

[0288] Feature fusion and optimization: The three layers of features are spliced ​​together through skip connections and assigned different weights through the attention mechanism to obtain the final 1024-dimensional geological feature vector.

[0289] Subsequently, the feature importance adaptive adjustment algorithm was applied to optimize the feature vector. After 12 rounds of iteration, the weight of fault-related features increased from the initial 0.18 to 0.47, the weight of lithologic interface features increased from 0.22 to 0.38, and the weight of single rock layer thickness features decreased from 0.32 to 0.14.

[0290] Feature encoding quality assessment: The t-SNE dimensionality reduction technique was used to map high-dimensional feature vectors into two-dimensional space. It was found that samples with similar geological conditions were clearly clustered together, and samples in different fault-affected areas formed clear classification boundaries.

[0291] Through feature importance analysis, it was found that the characteristics of fault intersection areas, lithologic mutation areas and coal seam thickness change areas contribute most to subsidence prediction, which is consistent with geomechanics theory.

[0292] Deep learning model training example:

[0293] During the model training phase, the following measures were taken to address the limited data in the mining area:

[0294] Construction of historical data set: Historical data of 42 coal mines with similar geological conditions were collected, including geological characteristics and subsidence parameter data of 286 mining areas.

[0295] The coal seams range in thickness from 1.2 to 6.8 meters, with burial depths from 120 to 520 meters. The degree of fault development varies, and mining techniques vary. The geological data for these mining areas are converted into feature vectors using the feature encoding network in step 1.

[0296] Data expansion: Physical model simulation and parameter transformation techniques were applied to generate an additional 1,420 sets of training samples based on the laws of geomechanics.

[0297] Implementation of innovative training strategies:

[0298] Multi-stage training: In the first stage, the parameters of specific branches are frozen, trained for 38 rounds, and the shared feature extraction layer is optimized; in the second stage, the shared layer is frozen and each parameter branch is trained separately for 24 rounds; in the third stage, the entire network is unfrozen and global fine-tuning is performed for 52 rounds.

[0299] Combined loss function: A combined loss function including prediction loss, constraint loss and regularization term is designed:

[0300]

[0301] Among them, L total Represents the total loss function, which is used to guide the model training process; represents the weighted sum of all settlement parameter prediction losses, i represents different settlement parameter indexes, such as subsidence coefficient, horizontal movement coefficient, influence angle, etc.; w′ i The weight coefficient for predicting the loss of subsidence parameters is used to balance the importance of different parameters in the total loss; is the predicted loss of each subsidence parameter, which is used to measure the difference between the predicted value and the true value; L is a constraint loss term used to penalize prediction results that violate physical laws and ensure that the prediction parameters satisfy the geomechanical relationship; reg is the model parameter regularization term, used to prevent the model from overfitting; c is the constraint weight coefficient, which controls the influence of the physical constraint term on the total loss; r is the regularization coefficient, which controls the influence of the regularization term on the total loss.

[0302] The prediction loss uses mean square error, the constraint loss uses L1 norm to measure the degree of deviation of the physical relationship between parameters, and the regularization term uses L2 regularization to prevent overfitting.

[0303] The initial weight is set to: w q =0.4, w b =0.3, w β =0.3,λ c =0.2,λ r =0.01.

[0304] Dynamic weight adjustment: After every five rounds of training, the loss weights are dynamically adjusted based on the prediction errors of each parameter on the validation set. During training, the weight of the sinking coefficient was gradually adjusted from an initial 0.4 to 0.25, while the weight of the influence angle increased from 0.3 to 0.45, reflecting the difficulty of predicting the influence angle.

[0305] Model evaluation and optimization:

[0306] Model performance evaluation: On the test set, the average relative error of the sinking coefficient was 4.2%, the horizontal movement coefficient was 5.7%, and the influence angle was 6.8%. The overall prediction accuracy was significantly better than the traditional method (the average relative error of the traditional method was 18.5%).

[0307] Model optimization: Through model pruning guided by sensitivity analysis, 128 connections with contributions below 1% were removed, reducing the number of model parameters by 21.3% and increasing inference speed by 32.5%, while maintaining prediction accuracy loss of less than 0.5%.

[0308] Improved ensemble learning: Five models with slightly different structures were trained, and the prediction results were fused using a weighted average method, which further improved the final prediction accuracy by 2.8 percentage points.

[0309] Application examples of transfer learning framework:

[0310] Given the limited monitoring data in this coal mining area (only 23 monitoring points), we applied the transfer learning framework to transfer the model knowledge trained in step 2 to this mining area to achieve accurate parameter inversion:

[0311] Domain difference analysis and feature alignment:

[0312] Domain difference analysis: The maximum average difference between the geological characteristic vectors of the source domain (historical mining area) and the target domain (this mining area) is 0.382, indicating that there are obvious differences between the two domains.

[0313] t-SNE visualization shows that the source and target area samples form two clearly separated clusters in the feature space, among which the differences in fault characteristics and lithology distribution characteristics are particularly significant.

[0314] Feature alignment implementation: A domain-invariant feature extractor was constructed, which consists of three fully connected layers with hidden layer dimensions of 512 and 256. A two-branch discriminator network was also designed to distinguish which domain the features come from.

[0315] During the adversarial training process, the feature extractor parameters are updated using a gradient reversal layer with a learning rate of 0.0005, an Adam optimizer, and a batch size of 64.

[0316] After 58 rounds of iterative training, the MMD values ​​of the two domains were reduced to 0.074. t-SNE visualization showed that the samples of the two domains had been fully mixed, indicating that the domain differences were effectively eliminated.

[0317] Hierarchical transfer learning model construction:

[0318] Bottom-layer network processing: The first two layers of the deep learning model's feature extraction network are completely frozen, retaining the basic geological feature extraction capabilities learned from the source domain. This part of the network mainly extracts common features such as stratum thickness changes and lithology distribution, and has high portability between different mining areas.

[0319] Intermediate layer network adjustment: Light fine-tuning is applied to the 3rd and 4th layer feature extraction networks, the learning rate is set to 1 / 10 of the base learning rate, and the adaptive batch normalization technology is introduced.

[0320] Specifically, the mean and variance of the target domain data are calculated to replace the batch normalization statistics of the source domain, so that the feature distribution adapts to the target domain while retaining the network weight parameters.

[0321] Retraining the top-level network: The parameter-specific branch network was completely retrained using data from 23 monitoring points in the target area, with a learning rate of 0.001 and 120 iterations. For the strongly coupled sinking coefficients and horizontal shift coefficients, a parameter sharing layer was designed to further improve data utilization efficiency.

[0322] Active learning monitoring point optimization layout:

[0323] Initial data analysis: Analysis of the data from the first 10 monitoring points revealed that the uncertainty of the prediction results was highest in the northeast and southwest of the mining area, and these areas also lacked representative monitoring points.

[0324] Application of hybrid sampling strategy: The hybrid sampling algorithm is applied to optimize the layout of the remaining 13 monitoring points. The uncertainty metric U(x) and representative metric R(x) are designed:

[0325] U(x) is the prediction variance based on the model ensemble, i.e., the standard deviation of the predictions of the five models at position x;

[0326] R(x) is measured by calculating the Euclidean distance between the position x and the nearest monitoring point in the feature space.

[0327] Dynamic adjustment of balance factor:

[0328] At the initial stage of monitoring point arrangement (11th to 15th points), the balance factor λ balance It is set to 0.35, focusing on improving the representativeness of the distribution of monitoring points;

[0329] As the number of monitoring points increases (from the 16th to the 20th point), λ balance Adjusted to 0.6, focusing more on high uncertainty areas;

[0330] In the final stage (points 21 to 23), λ balance Improved to 0.8, focusing on accurately determining the subsidence parameters in key areas.

[0331] Monitoring point layout results: The 23 monitoring points arranged using this strategy provide the same amount of information as the 42 monitoring points provided by the traditional equidistant layout method, improving monitoring efficiency by 82.6%. In particular, in areas with dense faults and areas with drastic lithologic variations, prediction accuracy improved by 24.3% compared to the equidistant layout.

[0332] Incremental learning and continuous updating:

[0333] Application of knowledge distillation: As monitoring data accumulates (from an initial 10 points to a final 23 points), knowledge distillation is used to continuously update the model. Specifically, we set the temperature parameter T = 3.0 and the distillation weight α = 0.7, so that the new model can learn the label information of the newly added monitoring points while retaining the original model's predictive ability for other areas.

[0334] Elastic Weight Consolidation: The EWC algorithm is introduced to prevent catastrophic forgetting and impose regularization constraints on important parameters. First, the parameter importance matrix (Fisher information matrix) is calculated, and then a penalty is imposed on changes in important parameters, with a penalty coefficient set to 2000. This allows the model to adapt to new monitoring point data while maintaining its predictive power in areas without monitoring data.

[0335] Model integration and update: As monitoring data increases, new models are continuously trained and integrated with the original model. In the initial stage, the weight of the original model is 0.8 and the weight of the new model is 0.2;

[0336] As the number of new monitoring point data increases to 15, the weights of the two are adjusted to 0.5:0.5;

[0337] When the number of monitoring points reaches 23, the weight of the original model drops to 0.2, and the weight of the new model increases to 0.8.

[0338] Parameter sensitivity analysis model example:

[0339] In order to optimize the selection of subsidence prediction parameters, a comprehensive sensitivity analysis of the subsidence prediction model for this coal mining area was conducted:

[0340] Local sensitivity analysis implementation:

[0341] Single parameter perturbation analysis: For the five key parameters, including subsidence coefficient q, horizontal movement coefficient b, influence angle β, inflection point offset distance S and mining influence propagation angle θ, perturbations of ±5%, ±10% and ±15% were performed respectively to observe the changes in the subsidence prediction results.

[0342] The results show that the sensitivity coefficient of the subsidence coefficient is the highest, reaching 0.83, followed by the influence angle (0.65) and the horizontal movement coefficient (0.51). The sensitivity coefficient of the inflection point offset is 0.42, the mining influence propagation angle is 0.38, and the displacement angle has the lowest sensitivity, only 0.12.

[0343] Application of gradient analysis: Automatic differentiation technology is used to calculate the gradient of the subsidence prediction results with respect to various parameters. The normalized gradient index shows that in areas with dense faults, the gradient value of the fault influence coefficient increases significantly, from an average of 0.22 to 0.58;

[0344] In the edge of the subsidence basin, the gradient values ​​of the inflection point offset S and the mining impact propagation angle θ increased significantly, reaching 0.47 and 0.53 respectively, indicating that these two parameters have a decisive influence on the prediction of the subsidence boundary.

[0345] In the lithologic homogeneous area, the gradient value of the lithologic correction coefficient is only 0.13, indicating that the parameter sensitivity has obvious spatial variation characteristics.

[0346] Application of global sensitivity analysis technology:

[0347] Implementation of the Sobol method: 10,000 parameter samples were generated and a global sensitivity analysis was performed using a Latin hypercube sampling design. The first-order sensitivity index and the total sensitivity index were calculated, and the results showed that:

[0348] First-order sensitivity index of sinking coefficient The total sensitivity index is 0.42. It is 0.65, indicating that it has both a strong direct effect and an interaction effect with other parameters;

[0349] Total sensitivity index of impact angle (0.58) is much higher than its first-order sensitivity index (0.31), indicating that it mainly affects the subsidence prediction results through interaction with other parameters;

[0350] First-order sensitivity index of inflection point offset S The total sensitivity index is 0.23. is 0.48, indicating that it has an important influence on the morphology of the settlement curve;

[0351] The first-order sensitivity index of mining impact on propagation angle θ The total sensitivity index is 0.25. It is 0.45, which is crucial for predicting the boundary of the subsidence impact range;

[0352] The second-order interaction sensitivity index of the sinking coefficient and the influence angle is found It reaches 0.15, indicating that there is a strong coupling relationship between the two. In addition, the second-order interaction index of the inflection point offset S and the mining impact propagation angle θ is 0.12, which also shows a significant coupling effect.

[0353] Application of the Morris screening method: 20 parameter trajectories were designed, each containing 10 levels, and the basic effect of each parameter was calculated.

[0354] The results show that the fault influence coefficient has not only a high average basic effect (1.42) but also a high standard deviation (0.78), indicating that it has not only a significant impact but also a strong nonlinear effect.

[0355] The inflection point offset S and mining impact propagation angle θ show higher basic effects (1.25 and 1.18, respectively) and standard deviations (0.65 and 0.61, respectively) in the boundary area, indicating that they have significant nonlinear effects and interactions in boundary prediction.

[0356] Parameter importance assessment based on machine learning:

[0357] Random Forest Importance Analysis: A random forest model with 500 trees was trained and parameter importance was assessed using feature permutation. The results showed that the importance of the subsidence coefficient was 100 (normalized), the influence angle was 78, and the horizontal shift coefficient was 65 for predicting maximum subsidence.

[0358] When predicting the location of the subsidence boundary, the importance of the influence angle rises to 92, exceeding the 87 of the subsidence coefficient. The importance score of the inflection point offset distance S is 85, and the importance score of the mining influence propagation angle θ is 82, highlighting the key role of these two parameters in boundary prediction.

[0359] SHAP value analysis: Calculate the SHAP value of each parameter for subsidence prediction at different locations. In the center of the mining area, the average SHAP value of the subsidence coefficient is 0.58; at the edge of the subsidence basin, the average SHAP value of the influence angle reaches 0.67, becoming the most critical parameter;

[0360] In the area extending outward from the subsidence boundary, the SHAP values ​​of the inflection point offset S and the mining impact propagation angle θ reach 0.65 and 0.62, respectively, becoming the decisive factors for the prediction in this area.

[0361] Especially in the fault intersection area, the SHAP value of the fault influence coefficient shows an obvious peak (0.45), indicating that the importance of fault parameters is significantly increased under complex geological conditions.

[0362] Parameter sensitivity visualization and optimization guidance:

[0363] Multidimensional sensitivity visualization: Generates parameter sensitivity heat maps, radar charts, and interaction matrices to intuitively display the importance and interaction relationships of each parameter.

[0364] The heat map shows a distinct spatial distribution of parameter sensitivities within the mining area, particularly with a "jump" in parameter sensitivity on either side of the fault. At the edge of the subsidence basin, the sensitivity to the inflection point offset S and the mining impact propagation angle θ is significantly higher than in other areas, forming a significant "edge high-sensitivity zone."

[0365] Parameter optimization strategy formulation: Based on the sensitivity analysis results, a zoning optimization strategy was proposed for the inversion of subsidence parameters in this mining area:

[0366] The mining area is divided into five parameter zones, each zone has similar parameter sensitivity characteristics;

[0367] Fine-grained step sizes (0.01 and 0.2°) were used to finely optimize highly sensitive parameters (sinking coefficient, influence angle);

[0368] The centering sensitivity parameters (horizontal movement coefficient, inflection point offset S, mining influence propagation angle θ) were optimized using a medium step size (inflection point offset step size of 0.5m, mining influence propagation angle step size of 0.5°);

[0369] For less sensitive parameters (shift angle), a coarse-grained step size or a fixed typical value is used.

[0370] Application of adaptive parameter adjustment mechanism: Develop an iterative optimization algorithm based on sensitivity and dynamically adjust the parameter search strategy:

[0371] Initial parameter estimation: preliminary setting of parameter values ​​based on geological conditions;

[0372] Sensitivity analysis: Calculate the sensitivity index of each parameter under the current parameter settings;

[0373] Parameter optimization: Determine the parameter optimization priority and search step size based on the sensitivity index;

[0374] Iterative process: 5 rounds of iterative optimization were performed until the subsidence prediction error dropped below the target threshold (5%).

[0375] Through the parameter optimization process driven by this sensitivity analysis, the optimal subsidence parameter combination for the mining area was finally determined: the subsidence coefficient q = 0.78 (corrected to 0.85 in the fault-affected area), the horizontal movement coefficient b = 0.32, the influence angle β varies with the lithology and fault distribution, and ranges from 65° to 78° in different areas, the inflection point offset S is 0.15H (H is the mining depth) in the flat area of ​​the coal seam, and 0.18H to 0.22H in the inclined area, and the mining influence propagation angle θ is 72°-78° in the fault-sparse area and decreases to 65°-70° in the fault-dense area.

[0376] Traditional methods require nearly 500 parameter combination evaluations, but the sensitivity analysis-driven optimization strategy only requires 72 evaluations to find the optimal parameter combination, increasing computational efficiency by nearly 7 times.

[0377] Technical effect verification:

[0378] After this implementation method was applied to this complex working face mine, it was verified through actual monitoring data and achieved significant technical results. The following is the verification data of two key technical effects:

[0379] Subsidence prediction accuracy is significantly improved:

[0380] To verify the prediction accuracy of this method, 32 test points were set up in the mining area (different from the 23 monitoring points used in training). The differences between the prediction results and the actual monitoring values ​​of the traditional method and this method were compared. The statistical results are as follows:

[0381] Comparison of prediction accuracy:

[0382] Traditional empirical parameter method: the average relative error is 28.7%, and the maximum relative error is 52.3%;

[0383] Traditional analogy method: The average relative error is 21.5%, and the maximum relative error is 36.8%.

[0384] The method of this application: the average relative error is only 6.2%, and the maximum relative error is 12.7%.

[0385] Prediction accuracy in areas with different geological conditions:

[0386] Normal geological area: the error of the traditional method is 15.6%, and the error of this method is 4.8%;

[0387] Fault-affected area: the traditional method has an error of 38.2%, while the present method has an error of 8.9%;

[0388] In the area of ​​lithologic variation, the error of the traditional method is 27.4%, while the error of this method is 6.5%;

[0389] In the overlapping area of ​​multiple working surfaces, the traditional method has an error of 42.1%, while the present method has an error of 9.2%.

[0390] Prediction accuracy of key parameters of surface subsidence:

[0391] Maximum subsidence value: the error of the traditional method is 22.3%, and the error of this method is 5.3%;

[0392] Maximum horizontal shift: the error of the traditional method is 25.7%, and the error of this method is 7.1%;

[0393] Subsidence basin extent: the traditional method has an error of 19.8%, while the present method has an error of 4.2%;

[0394] Maximum tilt deformation: the error of the traditional method is 32.5%, and the error of this method is 8.5%;

[0395] Inflection point offset S: The error of the traditional method is 29.3%, while the error of this method is 6.8%;

[0396] Mining-affected propagation angle θ: the traditional method has an error of 26.4%, while the present method has an error of 7.2%.

[0397] Prediction accuracy of different subsidence area locations:

[0398] Central area: The average error of the traditional method is 18.7%, and the average error of this method is 4.5%;

[0399] Edge transition area: the average error of the traditional method is 27.2%, and the average error of this method is 6.1%;

[0400] Inflection point region (the influence area of ​​the key parameter S): the average error of the traditional method is 31.5%, and the average error of this method is 7.8%;

[0401] Boundary extension area (mainly affected by θ): the average error of the traditional method is 33.2%, and the average error of this method is 8.3%.

[0402] The above data show that this method has significant advantages under different geological conditions and subsidence areas, especially in areas with complex geological conditions and the edge / inflection point of subsidence basins. The prediction accuracy is more significantly improved, with an average prediction accuracy increased by 76.8%, providing a reliable basis for safe mining decisions in mines.

[0403] In traditional methods, it is often difficult to accurately determine the inflection point offset distance S and the mining impact propagation angle θ. This method significantly improves the prediction accuracy of these two parameters by combining geological feature coding and parameter sensitivity analysis, providing strong support for the accurate prediction of subsidence boundaries.

[0404] Monitoring data requirements are significantly reduced:

[0405] In order to verify the performance of this method under limited data conditions, a prediction accuracy comparison experiment with different numbers of monitoring points was designed:

[0406] Comparison of prediction accuracy under different numbers of monitoring points:

[0407] 10 monitoring points: the average error of the traditional method is 48.5%, and the average error of this method is 14.7%;

[0408] 15 monitoring points: the average error of the traditional method is 36.2%, and the average error of this method is 10.3%;

[0409] 20 monitoring points: the average error of the traditional method is 27.8%, and the average error of this method is 7.8%;

[0410] 23 monitoring points: the average error of the traditional method is 22.3%, and the average error of this method is 6.2%;

[0411] 40 monitoring points: the average error of the traditional method is 15.2%, and the average error of this method is 5.8%.

[0412] Comparison of the number of monitoring points required to achieve the same prediction accuracy:

[0413] Achieve 15% prediction error: The traditional method requires 38 monitoring points, while this method only requires 10 monitoring points;

[0414] Achieve 10% prediction error: The traditional method requires 47 monitoring points, while this method only requires 15 monitoring points;

[0415] Achieve 7% prediction error: The traditional method requires 60 monitoring points, while this method only requires 20 monitoring points.

[0416] The number of monitoring points required for prediction of key boundary parameters (inflection point offset S and mining impact propagation angle θ) is as follows:

[0417] The prediction accuracy of the inflection point offset S reaches 10%: the traditional method requires 52 monitoring points, while this method only requires 18 monitoring points;

[0418] The prediction accuracy of the mining impact propagation angle θ reaches 10%: the traditional method requires 48 monitoring points, while this method only requires 17 monitoring points;

[0419] The two parameters can achieve 10% accuracy at the same time: the traditional method requires 58 monitoring points, while this method only requires 20 monitoring points.

[0420] The impact of different monitoring point layout strategies on parameter prediction accuracy:

[0421] Traditional equidistant arrangement: with 23 monitoring points, the inflection point offset distance S has an error of 18.5%, and the mining-affected propagation angle θ has an error of 17.2%;

[0422] The mixed sampling arrangement of this method: when there are 23 monitoring points, the error of the inflection point offset distance S is 8.6%, and the error of the mining-affected propagation angle θ is 7.8%;

[0423] Improvement effect: With the same number of monitoring points, the prediction accuracy of key boundary parameters increased by 53.5% and 54.7%.

[0424] Monitoring cost savings analysis:

[0425] Construction and maintenance cost of a single monitoring point: approximately RMB 45,000 per year;

[0426] The monitoring cost required to achieve 10% prediction accuracy using traditional methods is: 47×4.5=2.115 million yuan / year;

[0427] The monitoring cost required to achieve 10% prediction accuracy of this method is: 15×4.5=675,000 yuan / year;

[0428] Monitoring cost savings: 1.44 million yuan / year, a savings rate of 68.1%;

[0429] The cost of accurate boundary prediction (including S and θ parameters) is: 2.61 million yuan / year for the traditional method, 900,000 yuan / year for this method, with a savings rate of 65.5%.

[0430] The data above demonstrates that this method significantly reduces the demand for monitoring data through a transfer learning framework and active learning strategies for optimizing monitoring point placement. Even with a limited number of monitoring points, it maintains high prediction accuracy, achieving the goal of "less monitoring, higher accuracy" and significantly reducing the cost of mining subsidence monitoring.

[0431] In particular, for the two parameters of inflection point offset distance S and mining impact propagation angle θ, which are difficult to accurately determine in traditional methods, this method greatly improves monitoring efficiency through intelligent boundary sensitivity analysis and feature learning, making accurate prediction of subsidence boundaries more economical and feasible.

[0432] like Figures 2 to 11 As shown, they are respectively the blocks before merging; the blocks after merging; the comparison of a small number of blocks before and after merging; the comparison of another small number of blocks before and after merging; the prediction results of the complex working face settlement system; the prediction plane of the complex working face settlement system; the prediction contour map of the complex working face settlement system; the prediction curvature result map of the complex working face settlement system; the prediction curvature plane map of the complex working face settlement system; the two-dimensional contour map of the prediction curvature of the complex working face settlement system.

[0433] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A method for predicting mining subsidence in a complex working face mine, characterized in that: The steps include: Collect and standardize the original geological information of the mining area, build a multi-level geological feature coding network, convert the geological information into high-dimensional feature vectors, and integrate local, medium-range and global geological features; Combine high-dimensional feature vectors with historical mining data to train a deep learning model, establish a mapping relationship between geological features and subsidence parameters, integrate physical constraints to ensure the rationality of the prediction results, and use a transfer learning framework to transfer the trained model knowledge to the target mining area; The model migrated to the target mining area is fine-tuned using limited field monitoring data to achieve accurate parameter inversion. Based on the current model prediction results, an active learning strategy is used to optimize the layout of subsequent monitoring points, improve the representativeness and information content of the newly added monitoring data, and combine parameter sensitivity analysis to optimize the selection and adjustment of subsidence parameters. For the optimized subsidence parameters and models, through the incremental learning mechanism, as new monitoring data continues to accumulate, the model is continuously updated and optimized to improve the long-term adaptability and prediction accuracy of the system, and ultimately improve the overall accuracy of subsidence prediction.

2. The method for predicting mining subsidence in a complex working face mine according to claim 1, characterized in that: The geological feature encoding network includes a three-dimensional convolutional neural network, a graph convolutional network and a self-attention mechanism, which are used to extract local, mid-level and global geological features respectively, and generate a comprehensive feature vector through a feature fusion mechanism.

3. The method for predicting mining subsidence in a complex working face mine according to claim 1, characterized in that: The deep learning model adopts a multi-task learning structure, which can simultaneously predict multiple subsidence parameters and improve the accuracy and physical rationality of the prediction through parameter sharing and physical constraints.

4. The method for predicting mining subsidence in a complex working face mine according to claim 1, characterized in that: The transfer learning framework includes domain difference analysis, feature alignment, hierarchical transfer and active learning monitoring point optimization layout, which can achieve high-precision parameter inversion when monitoring data is limited.

5. The method for predicting mining subsidence in a complex working face mine according to claim 1, characterized in that: The active learning strategy combines uncertainty sampling with representative sampling, dynamically optimizes the spatial distribution of monitoring points through comprehensive scoring, and improves monitoring efficiency and prediction accuracy.

6. The method for predicting mining subsidence in a complex working face mine according to claim 1, characterized in that: Parameter sensitivity analysis includes local sensitivity analysis, global sensitivity analysis and parameter importance assessment based on machine learning, which is used to guide the optimization of key parameters and model adjustment.

7. The method for predicting mining subsidence in a complex working face mine according to claim 1, characterized in that: The incremental learning mechanism includes knowledge distillation, parameter regularization and model integration update, which can continuously optimize model performance and prevent catastrophic forgetting as monitoring data continues to accumulate.

8. The method for predicting mining subsidence in a complex working face mine according to claim 1, characterized in that: It also includes an adaptive processing mechanism for heterogeneous geological structures, adopts recursive feature decomposition and dynamic boundary condition identification algorithms, and automatically adjusts the calculation accuracy and solution strategy according to the geological complexity of different areas, so that the system can maintain stability in discontinuous geological environments.

9. The method for predicting mining subsidence in a complex working face mine according to claim 1, characterized in that: It also includes time-space decoupling algorithms and dynamic influence function construction, which separates time series from spatial influences through four-dimensional tensor decomposition, and combines nonlinear time factor correction to solve the problem of subsidence evolution prediction under different mining sequences of multiple working faces, and realizes rapid evaluation of different mining sequence scenarios.

10. A complex working face mining subsidence prediction system for executing a complex working face mining subsidence prediction method according to any one of claims 1 to 9, characterized in that: include: Geological data processing module, used to collect and standardize geological information and construct feature coding network; Subsidence parameter mapping module, used to train deep learning models to establish mapping relationships between geological features and subsidence parameters; Model fine-tuning and optimization module, used to fine-tune the model and optimize the monitoring point layout using limited monitoring data; The incremental learning update module is used to continuously update the optimization model as new data accumulates. The modules work together to achieve high-precision prediction of subsidence in complex mining areas.

Images