An underground mine goaf deformation field high-precision monitoring and early warning method and system

Abstract

This invention relates to the field of mine safety monitoring technology, specifically to a high-precision monitoring and early warning method and system for deformation fields in underground mine goaf areas. The method includes the following steps: integrating multi-source data features to obtain preliminary cross-modal fusion features; based on environmental interference compensation features, the preliminary cross-modal fusion features, and the multi-source data features, obtaining accurate multi-source fusion features after interference compensation through attention-weighted fusion and feature filtering optimization; combining the temporal deformation trend features and spatial distribution features of the goaf area to extract spatiotemporal aggregated deformation features; introducing dynamic deformation baseline features adapted to real-time operating conditions, the accurate multi-source fusion features, and the spatiotemporal aggregated deformation features into a Kalman filter to construct an adaptive early warning threshold for monitoring and early warning. This invention is suitable for the harsh environment of underground mining operations, such as high dust, high water mist, and high vibration, achieving sub-millimeter-level precision, all-weather, continuous automated monitoring of the goaf deformation field.

Description

Technical Field

[0001] This invention relates to the field of mine safety monitoring technology, specifically to a high-precision monitoring and early warning method and system for deformation fields in underground mine goaf areas. Background Technology

[0002] Goaf areas in underground mines are underground spaces formed after mining operations. Their rock walls and supporting structures are prone to deformation due to mining disturbances, load changes, and environmental erosion. If monitoring is not timely or early warnings are inaccurate, major safety accidents such as collapses can easily occur. Currently, goaf deformation monitoring technology mainly relies on single or multi-source equipment combinations such as lidar, inertial navigation (INS), visual sensors, and strain sensors. However, practical applications face the following core challenges:

[0003] Poor adaptability to harsh environments: Underground mining areas are generally characterized by high dust, high water mist, and high vibration, resulting in high noise in the original monitoring data, easy distortion of optical data, and easy drift of sensors. Traditional fixed threshold denoising and static standardization methods are difficult to effectively process such non-stationary interference data, and the data preprocessing accuracy is insufficient.

[0004] Insufficient fusion of multi-source data: Existing technologies often use simple splicing or weighted summation to fuse multimodal data, failing to fully explore the local details and global correlations of each modality. This results in the limitations of single-modal data and insufficient complementarity, making it difficult to form a comprehensive expression of deformation features.

[0005] Insufficient mining of temporal and spatial features: The deformation of goaf has an inherent law of temporal accumulation and spatial transmission, but traditional temporal models are prone to gradient vanishing problem and cannot effectively capture long-term deformation trends; spatial interpolation does not consider the reliability differences of monitoring points in harsh environments, and spatial clustering does not combine the regional characteristics of goaf deformation, resulting in distorted extraction of spatial distribution features.

[0006] Poor adaptability to dynamic working conditions: Traditional deformation baselines often use fixed values ​​or simple linear fitting, which cannot adapt to changes in dynamic working conditions such as mining progress and blasting vibration, and are prone to baseline deviation; warning thresholds are mostly fixed and are not dynamically adjusted in combination with real-time working conditions, resulting in a high rate of false and missed warnings.

[0007] Passive handling of environmental interference: Existing technologies mostly use simple filtering or ignore the impact of environmental interference, failing to achieve accurate quantification and active compensation of interference, making it difficult to meet the needs of sub-millimeter-level high-precision monitoring.

[0008] Therefore, designing a method that can achieve accurate data preprocessing, deep fusion of multi-source features, dynamic baseline adaptation, active interference compensation, and high-precision monitoring and early warning, in view of the harsh environment and dynamic working conditions of underground mine goaf areas, has become a key technical problem that urgently needs to be solved in the field of mine safety monitoring. Summary of the Invention

[0009] To address the shortcomings of existing methods and the needs of practical applications, and in order to solve the aforementioned problems, this invention provides a high-precision monitoring and early warning method for deformation fields in underground mine goaf areas, comprising the following steps:

[0010] Multi-source data features are integrated to obtain preliminary cross-modal fusion features. Based on environmental interference compensation features, the preliminary cross-modal fusion features, and the multi-source data features, attention-weighted fusion and feature selection optimization are used to obtain accurate multi-source fusion features after interference compensation. Combined with the temporal deformation trend features and spatial distribution features of the goaf, spatiotemporal aggregated deformation features are extracted. The dynamic deformation baseline features adapted to real-time working conditions, the accurate multi-source fusion features, and the spatiotemporal aggregated deformation features are introduced into Kalman filtering to construct an adaptive early warning threshold for working conditions to complete monitoring and early warning.

[0011] Optionally, the process of integrating multi-source data features to obtain preliminary cross-modal fusion features includes the following steps:

[0012] Extract local features of a single modality; based on the local features of the single modality, introduce a Transformer encoder to achieve global correlation fusion, and obtain preliminary cross-modal fusion features containing local details of each modality and global correlation.

[0013] Optionally, extracting the environmental disturbance compensation features includes the following steps:

[0014] An SVR interference quantification model is constructed; based on interference factor data, the influence of interference on monitoring data is quantified through the SVR interference quantification model, interference compensation coefficients are generated, and environmental interference compensation characteristics containing interference compensation coefficients of each monitoring data are obtained.

[0015] Optionally, the step of obtaining accurate multi-source fusion features after interference compensation based on environmental interference compensation features, the preliminary cross-modal fusion features, and the multi-source data features through attention-weighted fusion and feature selection optimization includes the following steps:

[0016] Based on the environmental interference compensation features, the cross-modal preliminary fusion features, and the multi-source data features, an attention mechanism is introduced to calculate the weights of each feature; L1 regularization features are used to filter and eliminate redundant features to achieve multi-feature fusion, thereby obtaining accurate multi-source fusion features after interference compensation.

[0017] Optionally, the step of extracting spatiotemporal aggregated deformation features by combining the temporal deformation trend characteristics and spatial distribution characteristics of the goaf includes the following steps:

[0018] By combining the temporal and spatial characteristics of the deformation trend and spatial distribution of the goaf, the intrinsic relationship between temporal and spatial features is captured. Residual connections are introduced to avoid information loss during feature aggregation. Based on the intrinsic relationship, deep aggregation of spatiotemporal features is achieved to obtain spatiotemporal aggregated deformation features that contain the intrinsic relationship between temporal trends and spatial distribution.

[0019] Optionally, the temporal deformation trend characteristics of the goaf are extracted, including the following steps:

[0020] A spatiotemporal attention mechanism is introduced to highlight deformation information at key time points; the gating unit of the LSTM network is optimized, and based on the deformation information, long-term deformation trend features containing deformation information of key time periods are extracted through the optimized LSTM network.

[0021] Optionally, extracting the spatial distribution characteristics of the goaf includes the following steps:

[0022] The Kriging interpolation algorithm is optimized by introducing a reliability weight for monitoring points. Based on the optimized Kriging interpolation algorithm, K-means++ spatial clustering is used to divide the key deformation regions and extract spatial distribution features.

[0023] Optionally, the step of introducing the dynamic deformation baseline features adapted to real-time operating conditions, the precise multi-source fusion features, and the spatiotemporal aggregated deformation features into a Kalman filter to construct an adaptive early warning threshold for operating conditions and complete monitoring and early warning includes the following steps:

[0024] Using the dynamic deformation baseline as a reference, and combining the precise multi-source fusion features and the spatiotemporal aggregated deformation features, deformation field reconstruction is achieved through Kalman filtering; based on the statistical characteristics of monitoring data and operating condition thresholds, an adaptive early warning threshold is constructed, and by combining the reconstruction results and the adaptive early warning threshold, deformation early warning with sub-millimeter accuracy is achieved.

[0025] Optionally, analyzing the dynamic deformation baseline characteristics adapted to real-time operating conditions includes the following steps:

[0026] Quantify the operating condition factors; use the operating condition factors as constraints, introduce an improved Bayesian inference algorithm to construct a dynamic deformation baseline, and analyze the characteristics of the dynamic deformation baseline adapted to the real-time operating conditions.

[0027] Secondly, to efficiently execute the high-precision monitoring and early warning method for deformation fields in underground mine goafs provided by this invention, this invention also provides a high-precision monitoring and early warning system for deformation fields in underground mine goafs, including a processor, an input device, an output device, and a memory. The processor, input device, output device, and memory are interconnected. The memory stores a computer program containing program instructions. The processor is configured to call the program instructions to execute the high-precision monitoring and early warning method for deformation fields in underground mine goafs as described in the first aspect of this invention. The high-precision monitoring and early warning system for deformation fields in underground mine goafs of this invention has a compact structure and stable performance, and can stably execute the high-precision monitoring and early warning method for deformation fields in underground mine goafs provided by this invention, further improving the overall applicability and practical application capability of this invention.

[0028] Beneficial Effects: This invention obtains preliminary cross-modal fusion features by integrating multi-source data, and then obtains precise fusion features through attention-weighted fusion and filtering optimization by combining environmental interference compensation features. Finally, it extracts spatiotemporal aggregation features by fusing temporal deformation trends and spatial distribution features. These features, along with dynamic deformation baseline features adapted to real-time operating conditions, are introduced into a Kalman filter, and monitoring and early warning are completed by combining them with an adaptive early warning threshold based on operating conditions. This process effectively solves problems such as insufficient multi-source data fusion, inadequate spatiotemporal feature mining, and poor adaptability to dynamic operating conditions under harsh environments. It strengthens interference suppression and core feature representation capabilities, significantly improves the accuracy of goaf deformation monitoring and early warning, and provides reliable technical support for mine safety. Attached Figure Description

[0029] Figure 1 A flowchart of a high-precision monitoring and early warning method for deformation field in underground mine goaf provided by an embodiment of the present invention;

[0030] Figure 2 This is a framework diagram of a high-precision monitoring and early warning system for deformation fields in underground mine goaf areas, provided as an embodiment of the present invention. Detailed Implementation

[0031] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0032] Throughout this specification, references to an embodiment, example, or illustration mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, phrases appearing in various places throughout the specification, such as "in one embodiment," "in an embodiment," "an example," or "an illustration," do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in any suitable combination and / or sub-combination in one or more embodiments or examples. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0033] Please see Figure 1 To address the aforementioned problems, this invention provides a high-precision monitoring and early warning method for deformation fields in underground mine goaf areas, such as... Figure 1 As shown, in one embodiment, the method includes the following steps:

[0034] S1. Integrate features from multiple data sources to obtain preliminary cross-modal fusion features.

[0035] Considering the dispersed installation locations of monitoring equipment in underground goaf areas—such as lidar installed on the top of roadways, visual sensors installed on supporting columns, and INS integrated into mobile monitoring platforms—data spatial misalignment and temporal asynchrony issues are prone to occur. Therefore, it is necessary to first unify the spatial reference and complete temporal synchronization. For spatial alignment, a unified three-dimensional coordinate system for the underground goaf area is established, using the spatial coordinates of the INS positioning data as the reference. The origin is the center point of the goaf entrance, the horizontal direction is the X-axis, the vertical direction is the Y-axis, and the roadway extension direction is the Z-axis. For lidar point cloud data, the discrete point cloud coordinates are mapped to this unified coordinate system using the nearest neighbor interpolation method. During the interpolation process, the point cloud offset is corrected by incorporating the lidar's installation attitude parameters to ensure that the point cloud data is consistent with the actual spatial location of the goaf area. For visual sensor image pixel coordinates, the pixel coordinates are converted into three-dimensional spatial coordinates under the unified coordinate system through camera intrinsic parameter calibration and extrinsic parameter estimation, achieving spatial alignment between visual data and lidar and INS data. In terms of time synchronization, GPS timestamps are used as the core synchronization benchmark. Timestamp information of data from various devices is collected, and the time difference between data from different devices is calculated. Linear interpolation is used to fill in the time gap data. Data with missing timestamps is marked as invalid data and removed. Ultimately, spatiotemporal homogeneity of multi-source data such as LiDAR, INS, vision sensors, and strain sensors is achieved, that is, accurate matching of multi-dimensional data at the same time and in the same spatial location, laying the foundation for subsequent cross-modal feature fusion.

[0036] Because the output dimensions of various monitoring devices differ significantly, direct fusion would lead to an imbalance in feature weights, necessitating standardization. Furthermore, considering the susceptibility of sudden changes in the underground mining subsidence area environment, an adaptive strategy combining Z-score standardization and dynamic mean correction is adopted. First, the sliding window size is determined to be 100 data sets. For each data set x to be standardized, its corresponding moving average is calculated. The sliding standard deviation s_t is used. When a sudden environmental change is detected, the sliding window size is automatically adjusted temporarily to 50 groups to speed up the update of standardized parameters and ensure that the standardized results can adapt to changes in data distribution in a timely manner. When the environment returns to stability, the window size is automatically switched back to 100 groups. In addition, to avoid abnormal standardized results caused by an excessively small sliding standard deviation s_t, a lower limit threshold of 0.001 is set. When s_t is lower than this threshold, s_t is forced to be 0.001 to prevent data distortion. Through this adaptive standardization process, multi-source data is uniformly mapped to the range of [-1,1] to obtain standardized multi-source data features with consistent spatiotemporal alignment, including standardized LiDAR point cloud coordinate features, standardized INS attitude features, standardized strain features, and standardized visual image grayscale features.

[0037] Furthermore, the process of integrating multi-source data features to obtain preliminary cross-modal fusion features includes the following steps:

[0038] S11. Extract local features of a single modality.

[0039] To address the multi-source data interference problem caused by high dust and water mist in underground mining operations, a modality-specific network is employed for precise local feature extraction, avoiding the inclusion of invalid information, including:

[0040] LiDAR point cloud feature extraction: For the standardized LiDAR point cloud data, the first step is to preprocess it based on the point cloud reflection intensity and spatial density. After setting a reflection intensity threshold to remove invalid points, the farthest point sampling algorithm is used to select core sampling points. Then, the XConv convolutional layer is used to capture the spatial topological relationship of the local point cloud and generate local clustering features.

[0041] Visual image feature extraction: To address the issues of image blurring and edge weakening caused by high water mist, based on the U-Net network, three sets of edge enhancement convolution kernels are added to the encoding end to perform edge enhancement processing on the RGB three channels of the image, strengthening the edge contour information of the goaf rock wall and support structure; at the decoding end, an attention gating unit is introduced to suppress background noise features caused by water mist and focus on deformation-sensitive areas (such as rock wall cracks and support column connections), and finally extract the image edge and texture fusion features, effectively improving the feature recognition in blurred environments.

[0042] Strain data feature extraction: For time-series data collected by strain sensors, a 1D-CNN network is used to extract local fluctuation features. The network structure is designed as 2 convolutional layers + 1 pooling layer. The first convolutional layer has a kernel size of 3 and a number of 64 to capture short-term local fluctuations in strain data (such as instantaneous strain changes caused by blasting vibrations). The second convolutional layer has a kernel size of 5 and a number of 128 to capture medium-term local trends (such as continuous strain accumulation caused by mining operations). A max pooling layer (pooling size 2) is inserted in the middle to reduce dimensionality and retain key features. Finally, the strain time-series local fluctuation features are output, which are adapted to the high-frequency fluctuation characteristics of strain data under high vibration environment.

[0043] Since the local feature dimensions extracted from different modalities differ, direct fusion can lead to an imbalance in feature weights. Therefore, an adaptive fully connected layer is needed to unify the dimensions. First, a dedicated fully connected mapping branch is configured for each modality feature, with the mapping formula f_unif = W·f_local + b, where W is the adaptive weight matrix specific to each modality and b is the bias term. During training, the values ​​of W and b are adaptively learned through backpropagation, with the objective function being to minimize the distribution differences of the mapped features from each modality, ensuring that different modality features are mapped to the same space.

[0044] S12. Based on the single-modal local features, a Transformer encoder is introduced to achieve global correlation fusion, and cross-modal preliminary fusion features containing local details of each modality and global correlation are obtained.

[0045] The multimodal features, after being unified in dimensionality, are concatenated to form the initial input features. These are then fed into a 6-layer Transformer encoder, with each layer employing a multi-head attention mechanism. The attention weight calculation process is as follows: the input features are transformed into a query matrix Q, a key matrix K, and a value matrix V through three independent linear transformation layers. After splitting Q, K, and V according to the number of heads, the attention weights for each head are calculated. The attention outputs of each head are concatenated, integrated through a linear transformation layer, and then processed by residual connections and LayerNorm normalization. Finally, a feedforward neural network outputs the final cross-modal preliminary fused features.

[0046] This process can effectively capture the global correlation between the spatial features, visual edge features, and strain temporal features of lidar. For example, when lidar detects a point cloud shift in a certain area, it can use an attention mechanism to correlate the edge changes and strain data fluctuations in the visual image of that area, thereby achieving complementary enhancement of multimodal information.

[0047] S2. Based on the environmental interference compensation features, the cross-modal preliminary fusion features, and the multi-source data features, the accurate multi-source fusion features after interference compensation are obtained through attention-weighted fusion and feature selection optimization.

[0048] In this embodiment, extracting the environmental interference compensation features includes the following steps:

[0049] S211. Construct an SVR interference quantization model.

[0050] Interference factors include dust concentration, water mist visibility, vibration amplitude, and temperature and humidity. To address the differences in the value range of different interference factors, Min-Max standardization is used to map them to the range of [0,1] to obtain a standardized interference factor vector.

[0051] Furthermore, a high-quality training dataset is constructed, selecting monitoring data under different underground working conditions (normal mining, blasting operations, equipment shutdown) and different interference intensities (low / medium / high dust, water mist, vibration), while simultaneously collecting baseline data from an interference-free environment. The input is a standardized interference factor vector, and the output is the difference Δi between the baseline data from the interference-free environment and the current environmental monitoring data, i.e., Δi = baseline data value - interfered monitoring data value. Δi > 0 indicates that the interference causes the monitoring data to be too small, and Δi < 0 indicates that the interference causes the monitoring data to be too large. To address the problem of insufficient high-interference-intensity samples, the SMOTE oversampling algorithm is used to expand the high-interference samples, while random undersampling is performed on low-interference samples to ensure a balanced proportion of samples at each interference intensity level and avoid the model from learning from low-interference samples.

[0052] An SVR interference quantization model is constructed based on the support vector machine model. A hybrid kernel function combining radial basis function and multinomial kernel function is used to optimize the SVR model's kernel function, balancing local and global fitting performance. The hybrid kernel function satisfies the following:

[0053] The optimal values ​​of each parameter are determined by grid search combined with 5-fold cross-validation. In this example, the weight coefficient a=0.6, the radial basis kernel parameter σ=0.8, the polynomial kernel constant term c=2, and the polynomial kernel degree d=3.

[0054] To prevent overfitting, an L2 regularization term (regularization coefficient C=1.2) is introduced to constrain the model parameter complexity. An ε-insensitive loss function (ε=0.001, adapted to sub-millimeter-level monitoring accuracy requirements) is adopted to allow for small errors and improve the model's generalization ability. During the training of the SVR interference quantization model, gradient descent is used to optimize the objective function, which minimizes the mean square error between the predicted and actual interference effects.

[0055] S212. Based on the interference factor data, the influence of interference on the monitoring data is quantified through the SVR interference quantification model, and interference compensation coefficients are generated to obtain environmental interference compensation characteristics that include the interference compensation coefficients of each monitoring data.

[0056] Based on the prediction results of the SVR interference quantization model, a dynamic adaptive interference compensation coefficient is constructed.

[0057] First, the interference impact is analyzed. The standardized interference factor vectors collected in real time are input into the trained SVR interference quantification model, and the interference impact Δi corresponding to each monitoring data (LiDAR coordinates, strain values, INS attitude parameters) is output. The interference is then broken down by type to obtain the dust interference impact Δi1, water mist interference impact Δi2, vibration interference impact Δi3, and temperature and humidity interference impact Δi4. The impact weight of each interference factor on different monitoring data is then determined. For example, vibration interference has the greatest impact on INS attitude data, with a weight of 40%; water mist interference has the greatest impact on visual image grayscale data, with a weight of 35%.

[0058] Then, the factor compensation coefficients and the overall compensation coefficient are calculated. The factor compensation coefficients satisfy: ,in Let j be the amount of impact of the j-th type of interference on the i-th type of monitoring data. The real-time value of the i-th type of monitoring data is used to individually correct the impact of a certain type of interference; the comprehensive compensation coefficient k_i is the weighted sum of the compensation coefficients of each sub-factor, satisfying: k_i=Σ(wj× ), where wj is the weight coefficient of the j-th type of interference, determined based on the analytic hierarchy process, such as vibration interference w3=0.4, water mist interference w2=0.3, dust interference w1=0.2, temperature and humidity interference w4=0.1, used to correct the overall impact of all interferences.

[0059] In this embodiment, to avoid distortion of monitoring data due to abnormal compensation coefficients, a coefficient constraint range is set: k_i∈[0.8,1.2]. When the calculated k_i<0.8, k_i is forced to be 0.8; when k_i>1.2, k_i is forced to be 1.2. This prevents over- or under-compensation due to extreme interference effects, and finally generates an environmental interference compensation feature vector containing the sub-factor compensation coefficients and comprehensive compensation coefficients of each monitoring data.

[0060] Furthermore, the process of obtaining accurate multi-source fusion features after interference compensation based on environmental interference compensation features, the preliminary cross-modal fusion features, and the multi-source data features through attention-weighted fusion and feature selection optimization includes the following steps:

[0061] S221. Based on the environmental interference compensation features, the cross-modal preliminary fusion features, and the multi-source data features, an attention mechanism is introduced to calculate the weights of each feature.

[0062] Considering the heterogeneity of the three feature dimensions, a dedicated mapping branch is constructed to ensure that the core information of the features is preserved after mapping. An independent fully connected mapping branch is configured for each feature type. The branch structure is uniformly designed as a single hidden layer + output layer, and each branch embeds a Dropout layer. The specific configuration in the embodiment is as follows:

[0063] Multi-source data feature mapping branch: input dimension 192 (including standardized LiDAR, INS, strain, and visual features), hidden layer node number set to 224, output layer node number 256, activation function used is GELU;

[0064] Cross-modal preliminary fusion feature mapping branch: Input dimension 256 (cross-modal preliminary fusion features), only 1 output layer is needed to maintain dimensionality consistency, and a BatchNorm layer is embedded to standardize the feature distribution and accelerate subsequent fusion training;

[0065] Environmental interference compensation feature mapping branch: input dimension 64 (environmental interference compensation features), hidden layer node number set to 128, output layer node number 256, activation function used is GELU, and at the same time, 1×1 convolution layer is added at the input to compress redundant interference information and retain the core compensation coefficient features.

[0066] During the mapping process, the cosine similarity between the mapped features and the original features is ≥0.9 as a constraint. The weight matrix of each branch is optimized through backpropagation. The He initialization strategy is adopted to avoid gradient explosion in the early stage of training. Finally, three sets of 256-dimensional feature vectors are output to achieve a feature set to be fused with completely consistent dimensions, which provides a foundation for subsequent attention weighting.

[0067] Furthermore, to highlight the contribution of the dry environment interference compensation feature H and the cross-modal preliminary fusion feature B, and to suppress redundant information in the multi-source data feature A, a multi-head attention fusion module with initial weight guidance and adaptive adjustment of operating conditions is constructed.

[0068] First, the core structure of the module is determined: a 4-head attention mechanism is adopted, with each attention head having 64 dimensions. The MLP structure is designed as a 2-layer fully connected layer + GELU activation, with 512 nodes in the first layer and 256 nodes in the second layer, outputting a feature importance score. The logic for setting the initial weights is clear: the initial weight β_H0 for environmental interference compensation features is set to 0.4, the initial weight β_B0 for cross-modal preliminary fusion features is set to 0.4, and the initial weight β_A0 for multi-source data features is set to 0.2, and the initial weights satisfy the constraint Σβ_j=1.

[0069] Subsequently, an adaptive adjustment mechanism based on working conditions is introduced: combining the real-time working condition factor vector C (mining progress, load, etc.), it is mapped to a 3D weight correction coefficient Δβ=[Δβ_A,Δβ_B,Δβ_H] through a fully connected layer. The correction formula is β_j'=β_j+Δβ_j, and after correction, it is re-normalized to ensure Σβ_j'=1. For example, when blasting conditions are detected, Δβ_H is increased by 0.05~0.08 to strengthen the weight of environmental interference compensation features; when the mining progress is >70% (deformation risk increases), Δβ_B is increased by 0.05 to strengthen the global correlation information of cross-modal preliminary fusion features. Finally, the weight calculation is achieved through the formula β_j=Softmax(MLP(f_j)+β_j'), which retains the prior logic of the initial weights and adapts to the dynamic changes of underground working conditions, ensuring the rationality of weight allocation. f_j represents the j-th type of feature after mapping.

[0070] S222. Use L1 regularization to filter and remove redundant features to achieve multi-feature fusion and obtain accurate multi-source fusion features after interference compensation.

[0071] First, weighted fusion is performed. Based on the attention weights β_A, β_B, and β_H mentioned above, the initial fusion feature f_fusion = β_A·f_A + β_B·f_B + β_H·f_H is calculated. During the fusion process, residual connections f_fusion_init = f_fusion + concat(f_A,f_B,f_H)[:256] are introduced to preserve the core details of the original features and avoid information loss during the fusion process. Then, the feature distribution is normalized through the LayerNorm layer to stabilize the gradient of the subsequent screening process.

[0072] Subsequently, L1 regularized feature filtering is performed. The core objective is to eliminate redundant features. The regularization loss function is set as: L = ||f_fusion_init||1 + λ||W||1, where λ is the regularization coefficient and W is the weight matrix for weighted fusion. Specifically, the absolute values ​​of each dimension of f_fusion_init are calculated, retaining the feature dimensions with values ​​≥ λ and eliminating redundant dimensions with values ​​< λ. This yields accurate multi-source fusion features after interference compensation, ensuring that the features are both concise and accurate enough to support subsequent deformation monitoring.

[0073] S3. Combine the temporal deformation trend characteristics and spatial distribution characteristics of the goaf to extract the spatiotemporal aggregate deformation characteristics.

[0074] Extracting the temporal deformation trend characteristics of the goaf includes the following steps:

[0075] S311. Introduce a spatiotemporal attention mechanism to highlight deformation information at key time points.

[0076] Considering that critical working periods such as mining operations and blasting are high-incidence periods for goaf deformation, it is necessary to strengthen the feature weights of data during these periods through an attention mechanism, while suppressing redundant information during non-critical periods (such as equipment shutdown and ventilation operations). First, the core guiding vector g of the spatiotemporal attention module is constructed: based on the mining operation period markers (0 for non-operation, 1 for operation) provided in real time by the working condition monitoring system, it integrates quantitative working condition parameters such as mining intensity, peak blasting vibration, and support structure stress, and maps these parameters into a 128-dimensional guiding vector g through a fully connected layer, ensuring that the guiding vector can accurately characterize the degree of influence of different working conditions on deformation.

[0077] Subsequently, the attention weights for each time step are calculated. A dot product is performed on the feature vector s_t of time step t in the time series matrix with the guiding vector g, yielding the association value s_t·g, which characterizes the correlation between that time step and the key operating conditions. The association values ​​for all time steps are normalized using the Softmax function. The weight calculation formula is α_t = exp(s_t·g) / Σexp(s_i·g), increasing the weight α_t of time steps during key operating conditions (such as mining operations) to highlight key temporal information. To avoid excessively high weights for single time steps due to extreme conditions (such as high-intensity blasting), a weight upper limit threshold is set, and any excess is evenly distributed to adjacent time steps to prevent temporal feature distortion caused by weight imbalance. Finally, the feature vector of each time step in the time series matrix is ​​weighted and multiplied with its corresponding weight α_t to obtain a weighted time series matrix. This strengthens the deformation information during key operating conditions, providing a more targeted data foundation for subsequent LSTM feature extraction.

[0078] S312. Optimize the gating unit of the LSTM network, and extract long-term deformation trend features containing deformation information of key time periods through the optimized LSTM network based on the deformation information.

[0079] First, the LSTM network structure is determined. It is designed as a 3-layer stacked structure with 256 nodes in each hidden layer, 128-dimensional input layer (which is consistent with the feature dimension of the weighted time series matrix), and 128-dimensional output layer.

[0080] Subsequently, the activation function of the gated unit was optimized by replacing the Sigmoid function with Leaky-ReLU and setting the negative slope of Leaky-ReLU to 0.01 to ensure that there is still a weak gradient flow in the negative gradient region and to avoid the gradient from disappearing completely. The specific gating calculation process is as follows: The forget gate is responsible for filtering historical feature information, satisfying: f_t = LeakyReLU(W_f·[h_{t-1},x_t] + b_f), where W_f is a 128×256 dimensional weight matrix, initially initialized using Xavier to avoid gradient explosion in the early stages of training, h_{t-1} is the output of the hidden layer at the previous time step, containing historical deformation trend information, x_t is the weighted feature at the current time step, containing current deformation information, and b_f is the bias term, initially set to 0.1 to enhance the initial filtering capability of the forget gate. The value between 0 and 1 is output by f_t to control the retention ratio of historical features. During critical periods, f_t approaches 1 to retain more historical information; during non-critical periods, f_t approaches 0.5 to reduce redundant information interference. The input gate is responsible for updating the current feature information, satisfying: i_t = LeakyReLU(W_f·[h_{t-1},x_t] + b_f), where W_f is a 128×256 dimensional weight matrix, initially initialized using Xavier to avoid gradient explosion in the early stages of training, h_{t-1} is the output of the hidden layer at the previous time step, containing historical deformation trend information, x_t is the weighted feature at the current time step, containing current deformation information, and b_f is the bias term, initially set to 0.1 to enhance the initial filtering capability of the forget gate. The value between 0 and 1 is output by f_t to control the retention ratio of historical features. During critical periods, f_t approaches 1 to retain more historical information; during non-critical periods, f_t approaches 0.5 to reduce redundant information interference. kyReLU(W_i·[h_{t-1},x_t]+b_i), where W_i is a 128×256 dimensional weight matrix, b_i is a bias term, initially set to 0.1, and i_t outputs a value between 0 and 1, controlling the update ratio of the current feature. During periods of severe deformation, i_t approaches 1, prioritizing the update of the current deformation information. Combined with the cell state update formula C_t=f_t·C_{t-1}+i_t·tanh(W_C·[h_{t-1},x_t]+b_C), where W_C is the weight matrix, b_C is the bias term, and the tanh function normalizes the cell state to [-1,1], and the output gate formula o_t=LeakyReLU(W_o·[h_{t-1},x_t]+b_o), where W_o is the weight matrix, b_o is the bias term, the current hidden layer output h_t is obtained. The weighted time series matrix of multiple time steps is iteratively calculated by using an LSTM network, and the output of the last hidden layer is finally taken as the long-term deformation trend feature, which fully preserves the deformation fluctuation law and long-term deformation evolution trend of key working conditions.

[0081] Further, the spatial distribution characteristics of the goaf are extracted, including the following steps:

[0082] S321. Introduce a Kriging interpolation algorithm that optimizes the reliability weight of monitoring points.

[0083] High vibrations and electromagnetic interference in underground wells can easily lead to data distortion or failure at monitoring points. Therefore, a multi-dimensional reliability assessment system is constructed to accurately quantify the credibility of data from each monitoring point. First, core assessment indicators are determined. In addition to vibration amplitude and signal-to-noise ratio (SNR), key indicators such as sensor power supply stability, data transmission success rate, and historical data consistency are combined to form a 6-dimensional stability assessment vector. Then, the analytic hierarchy process (AHP) is used to determine the weights of each indicator. In this example, the weight of SNR is set to 0.4, vibration amplitude to 0.25, data transmission success rate to 0.15, historical data consistency to 0.1, and power supply stability to 0.1. After normalizing each indicator, the reliability weight ω_i is calculated by weighted summation, satisfying:

[0084] ω_i = 0.4 × SNR_norm + 0.25 × Vib_norm + 0.15 × Trans_norm + 0.1 × Consist_norm + 0.1 × Power_norm, where SNR_norm is the normalized signal-to-noise ratio, Vib_norm is the normalized vibration amplitude, Trans_norm is the normalized transmission success rate, Consist_norm is the normalized historical consistency, and Power_norm is the normalized power supply stability. The value of ω_i is strictly limited to [0,1]. When ω_i ≥ 0.8, it is considered a high-reliability monitoring point; when 0.5 ≤ ω_i < 0.8, it is considered a medium-reliability monitoring point; and when ω_i < 0.5, it is considered a low-reliability monitoring point. This evaluation system can effectively filter out problems such as sensor drift caused by high vibration and transmission distortion caused by electromagnetic interference, providing a fundamental guarantee for the accuracy of subsequent interpolation.

[0085] Kriging interpolation treats all known monitoring points as equally reliable, which can easily lead to interpolation distortion due to low-reliability data. Therefore, a reliability weight ω_i is introduced for optimization, enhancing the interpolation contribution of high-reliability points. First, data preprocessing is performed to remove extremely low-reliability points (ω_i < 0.3). For low-reliability data (0.3 ≤ ω_i < 0.5), the mean of three adjacent high-reliability points is used for correction to ensure the quality of the data input to the interpolation model.

[0086] Subsequently, an improved Kriging interpolation model was constructed. Based on the theory of regionalized variables, a spherical variogram was selected to fit the spatial correlation of known monitoring points. The variogram satisfies the following: Where h is the distance between monitoring points, (where C is the nugget value, C is the base value, and a is the range).

[0087] The parameters of the spherical variogram (gold nugget value) were obtained by fitting using the least squares method. After setting the sill value C and adjusting the range a), the objective function is to minimize the weighted interpolation error variance. The objective function satisfies:

[0088] Where γ is the value of the variogram. Let ω_i be the interpolation point and ω_i be the reliability weight of the monitoring point. The contribution of high reliability points is enhanced by weighting. Combined with the constraint Σλ_i(ω_i)=1 (to ensure unbiased interpolation), the Lagrangian function L=ΣΣλ_i(ω_i)·λ_j(ω_j)·γ(x_i,x_j)-2Σλ_i(ω_i)γ(x_i,x0)+μ(1-Σλ_i(ω_i)) is constructed, where μ is the Lagrange multiplier. The partial derivatives of L with respect to λ_i(ω_i) and μ are taken respectively and the partial derivatives are set to zero to obtain a linear equation system. Substituting the variogram parameters into the equation system, the weighted interpolation coefficients λ_i(ω_i) are finally obtained.

[0089] The final optimized interpolation formula is as follows: ,in Z(x_i) represents the estimated value of the point to be interpolated, and Z(x_i) represents the corrected known monitoring point data. The interpolation grid is set to 10cm×10cm to meet the sub-millimeter level monitoring accuracy requirements. The entire goaf area is divided into grids, and the interpolation value is calculated point by point to fill in the missing monitoring point data. For the grid edge areas, the extrapolation method is used in combination with the goaf boundary constraints to supplement the data, avoiding interpolation distortion caused by edge effects, and finally a uniformly distributed spatial data matrix is ​​obtained.

[0090] S322. Based on the optimized Kriging interpolation algorithm, K-means++ spatial clustering is used to divide the key deformation regions and realize the extraction of spatial distribution features.

[0091] First, the uniformly distributed spatial data matrix is ​​preprocessed, and PCA (Principal Component Analysis) is used for dimensionality reduction. The first 20 principal components are retained to reduce data redundancy while retaining core spatial information, thus avoiding the low clustering efficiency caused by high-dimensional data.

[0092] Subsequently, the K-means++ algorithm was used for spatial clustering. In this embodiment, the number of clusters was determined by maximizing the silhouette coefficient. Considering the common size of underground goaf areas, the number of clusters was divided into 5-8: goaf area <2000m² 2 The area is divided into 5-6 zones, with an area ≥2000m². 2 The time is divided into 7-8 zones, which can accurately cover key deformation areas such as the middle and edges of the rock wall, the area around the support pillars, and the area near the mining face.

[0093] After clustering, multi-dimensional statistical features are extracted for each cluster region: in addition to mean, variance, and extreme values, three key features are also included: spatial gradient (reflecting the rate of deformation change), centroid offset (reflecting the overall deformation trend of the region), and deformation duty cycle (reflecting the proportion of deformation points in the region). A 7-dimensional feature vector is generated for each region. Finally, the entropy weight method is used to determine the weight of each feature. In this example, the extreme values ​​and spatial gradient have the highest weights, both 0.25; the mean has a weight of 0.2; the variance and centroid offset have weights of 0.1; and the deformation duty cycle has a weight of 0.1. After weighted fusion, the spatial distribution features are obtained, which fully characterize the spatial heterogeneity of deformation in the goaf.

[0094] Furthermore, the extraction of spatiotemporal aggregated deformation features by combining the temporal deformation trend characteristics and spatial distribution characteristics of the goaf includes the following steps:

[0095] S331. Combining the temporal and spatial characteristics of the temporal deformation trend and spatial distribution of the goaf, capture the intrinsic relationship between temporal and spatial characteristics.

[0096] A channel-dimensional splicing method is used to sequentially concatenate temporal deformation trend features and spatial distribution features into an initial spatiotemporal feature f_init. Then, a spatiotemporal attention matrix is ​​constructed to accurately calculate the intrinsic correlation between each dimension of the temporal feature and each dimension of the spatial feature. Considering the existence of local anomaly correlations in downhole data (such as instantaneous spatiotemporal feature distortion caused by blasting), Pearson correlation coefficients and anomaly correlation filtering are used to capture the intrinsic correlation between temporal and spatial features. First, the correlation coefficient matrix between temporal feature c (c_i is the i-th temporal feature, corresponding to deformation fluctuations in key periods) and spatial feature d (d_j is the j-th spatial feature, corresponding to deformation differences in different regions) is calculated, with element γ_ij = Corr(c_i, d_j), where Corr is the Pearson correlation coefficient. To filter out invalid associations, such as false high associations caused by sensor drift, an association threshold γ_th = 0.1 is set. When |γ_ij| < γ_th, γ_ij is forced to = 0, retaining only the significant association dimension. At the same time, a spatiotemporal consistency check is introduced. If the coefficient of variation of a certain association value γ_ij with the adjacent 10 association values ​​is > 0.5, it is judged as an abnormal association, and the mean of the adjacent association values ​​is used to replace it to ensure the reliability of the attention matrix. Finally, an anti-interference spatiotemporal attention matrix is ​​obtained, which accurately describes the intrinsic association law of temporal deformation trend and spatial distribution.

[0097] S332. Introduce residual connections to avoid information loss during feature aggregation. Based on the intrinsic correlation, achieve deep aggregation of spatiotemporal features to obtain spatiotemporal aggregated deformation features that contain the intrinsic correlation between temporal trends and spatial distribution.

[0098] To achieve deep synergy of spatiotemporal features, weighted aggregation based on the attention matrix is ​​required, while anti-interference design and residual connections are introduced to enhance feature representation.

[0099] First, attention weights are normalized. Considering that abnormal correlations in the well may cause some γ_ij values ​​to be abnormally high, Softmax normalization is performed on each row of the attention matrix (corresponding to the correlation between a single temporal dimension and all spatial dimensions) to ensure that the sum of the weights in each row is 1, thus avoiding the aggregation process being dominated by a single abnormal correlation value.

[0100] Then, weighted aggregation is performed: the normalized attention weight γ'_ij is multiplied by the corresponding temporal dimension c_i and spatial dimension d_j in a triple product operation, that is, the single-element aggregation value is γ'_ij·c_i·d_j. Then, the summation of all elements is used to obtain the preliminary aggregated feature f_ts, which satisfies: f_ts=ΣΣγ'_ij·c_i·d_j.

[0101] To adapt to the strong coupling of downhole spatiotemporal characteristics, a robust weight adjustment mechanism is added. Elements contributing the top 10% of the aggregation process (i.e., elements with larger γ'_ij·c_i·d_j values, corresponding to core spatiotemporal associations) are given an additional 1.1 times enhancement weight to strengthen the expression of key spatiotemporal information; elements contributing the bottom 5% of the aggregation process are given a 0.8 times suppression weight to reduce interference from redundant information.

[0102] Next, residual connections are introduced. Considering that core information is easily lost during the aggregation of downhole spatiotemporal features (such as the spatiotemporal correlation of low-amplitude deformation being masked), the initial spatiotemporal feature f_init and the preliminary aggregated feature f_ts are added element-wise to obtain the enhanced spatiotemporal aggregated deformation feature f_ts_enh. This residual connection can directly retain the core details of the original spatiotemporal features (such as small spatial deformations in a certain region or instantaneous temporal fluctuations at a certain moment), while superimposing the aggregated correlation information, effectively avoiding the gradient vanishing problem and improving the feature's ability to represent complex downhole deformations.

[0103] In this embodiment, to adapt to the training requirements of the subsequent monitoring and early warning model, multiple optimization steps are used to eliminate abnormal feature distribution and redundant information, ensuring the stability and accuracy of spatiotemporal aggregate deformation features.

[0104] First, BatchNorm standardization is performed. To address the issue that downhole data distribution is prone to shifts in feature mean due to changes in working conditions, such as adjustments in mining intensity, the BatchNorm layer parameters are set to momentum = 0.9 and epsilon = 0.01. This standardizes the enhanced spatiotemporal aggregated feature f_ts_enh, eliminating the risk of gradient explosion caused by feature distribution shifts and accelerating subsequent model convergence.

[0105] Subsequently, L2 regularization was introduced to suppress overfitting. Considering that the downhole training samples may have local uneven distribution, such as fewer samples in high-interference conditions, the regularization coefficient λ=0.005 was set, and the loss function L=L_base+λ·Σ||w|| was applied. 2 The feature weights are constrained, with L_base as the base loss and w as the feature weights, to avoid the model overfitting to abnormal features of local samples in the well. The resulting spatiotemporal aggregated deformation features not only preserve the temporal trend and spatial distribution correlation of goaf deformation, but also have strong anti-interference and generalization capabilities.

[0106] S4. Introduce the dynamic deformation baseline features adapted to real-time operating conditions, the precise multi-source fusion features, and the spatiotemporal aggregated deformation features into the Kalman filter to construct an adaptive early warning threshold for operating conditions and complete monitoring and early warning.

[0107] In this embodiment, analyzing the dynamic deformation baseline characteristics adapted to real-time operating conditions includes the following steps:

[0108] S411, quantitative chemical condition factor.

[0109] Operating factors include mining progress, load changes, support structure status, and blasting vibration.

[0110] For mining progress data, the ratio of the mined volume to the total designed mining volume is calculated based on the GPS / INS fusion positioning data of mining equipment (such as fully mechanized mining machines and tunneling machines) and the mining boundaries in the goaf design drawings, as the mining progress condition factor.

[0111] For goaf load data, real-time pressure values ​​are collected by pressure sensors placed on the top and bottom plates of the goaf. Considering the uneven load distribution, the average value of the sensor data in each area is selected as the basic data and converted into pressure per unit area as the goaf load condition factor.

[0112] For the strain data of the support structure, the real-time strain value of the fiber optic strain sensor is collected and converted into a relative strain value as the condition factor of the support structure by the conversion formula of wavelength drift and strain ε=(λ-λ_0) / (K×λ_0), where λ is the real-time wavelength, λ_0 is the initial wavelength, and K is the sensor sensitivity coefficient.

[0113] For blasting vibration parameters, vibration signals are collected by blasting vibration sensors, and the peak acceleration of the signals is extracted as the blasting vibration condition factor.

[0114] After quantifying each factor, the quantification results are standardized by Min-Max (mapped to [0,1]) to construct a 4-dimensional working condition factor vector C=[mining volume ratio_norm, unit area pressure_norm, relative strain value_norm, peak vibration acceleration_norm]. Finally, the Local Anomaly Factor (LOF) algorithm is used to remove outlier vectors (LOF value > 1.5 is judged as anomaly) to avoid the influence of extreme working condition data (such as excessive vibration caused by blasting errors).

[0115] S412. Using the operating condition factor as a constraint, an improved Bayesian inference algorithm is introduced to construct a dynamic deformation baseline, and the characteristics of the dynamic deformation baseline adapted to the real-time operating condition are analyzed.

[0116] Bayesian inference does not fully incorporate the physical relationship between working conditions and deformation, which can easily lead to insufficient adaptability of the baseline model. By constraining working conditions factors, optimizing prior distributions, and improving the likelihood function, a dynamic baseline model that fits the actual downhole conditions is constructed.

[0117] Prior information and prior distribution selection: The standardized historical deformation data D is divided into stable and dynamic operating conditions. Stratified sampling is used to extract effective data as prior information. Considering the temporal continuity and randomness of the deformation data, a normal-inverse-gamma distribution is selected as the prior distribution P(θ|C). Its parameters are optimized by the operating condition factor C. For example, when the mining volume ratio is >50%, the variance of the prior distribution is increased to adapt to the large fluctuations in deformation under high mining intensity; the peak value of blasting vibration is >1 m / s. 2 At that time, the mean of the prior distribution is adjusted to match the instantaneous shift characteristics of the deformation after the blast.

[0118] Likelihood Function Construction: The likelihood function P(D|θ,C) is used to characterize the probability of historical deformation data D occurring given model parameters θ and operating condition factors C. To incorporate the physical laws of operating conditions and deformation, an improved Gaussian likelihood function is adopted, satisfying: d_t represents the historical deformation value at time t, and μ(θ,C) represents the mean deformation value under the working condition constraints. Let μ(θ,C) be the deformation variance under the constraints of the operating conditions; where μ(θ,C) is determined by a linear regression model combined with physical priors, i.e. , As the reference offset, These are the weighting coefficients for each operating condition factor. The quantified operating condition factors, weighting coefficients The sign of the symbol is constrained by physical laws, such as (Mining progress) Corresponding Increased mining progress leads to greater deformation.

[0119] Analysis of the core formula of the improved Bayesian inference model: Here, P(θ|D,C) is the posterior distribution, i.e., the probability distribution of the model parameter θ given historical data D and operating condition C, and ∝ indicates proportionality to. By incorporating information from historical deformation data into the model through the likelihood function P(D|θ,C), and simultaneously introducing constraint information from operating condition factors through the prior distribution P(θ|C), the solved model parameter θ can both conform to historical deformation patterns and adapt to the characteristics of current operating conditions, avoiding the limitations of traditional models that discuss deformation in isolation from operating conditions.

[0120] In addition, to improve adaptability to sudden underground working conditions, a dynamic weight coefficient η (η∈[0.8,1.2]) of the working condition factor is introduced. When a sudden change in working conditions is detected (such as blasting or a sudden increase in mining intensity), the value of η is increased (η=1.1~1.2) to strengthen the constraint effect of the working condition factor on the model parameters; when the working conditions are stable, the value of η is decreased (η=0.8~1.0) to weaken the constraint and retain the core patterns of historical data.

[0121] To achieve real-time tracking of operating condition changes by the baseline, the initial dynamic baseline is first solved. The Metropolis-Hastings (MH) sampling method in the Markov Chain Monte Carlo (MCMC) algorithm is used to solve for the parameter θ of the posterior distribution P(θ|D,C). Specifically, a Gaussian proposal distribution is used to generate candidate parameters θ*, with acceptance probabilities satisfying:

[0122] ,when When the random number U(0,1) is greater than or equal to the random number U(0,1), accept If a parameter is set as the new parameter, the original parameter is retained. The sampling convergence is determined by the R-hat convergence criterion (R-hat < 1.01). After convergence, the mean of the effective samples is taken as the optimal model parameter θ_opt. This is substituted into the baseline calculation formula B(x,t)=μ(θ_opt,C(t))+σ(θ_opt,C(t))×δ, where x is the coordinate of the monitoring point, t is the time, μ(·) is the mean deformation under the working condition constraint, σ(·) is the standard deviation of deformation under the working condition constraint, C(t) is the working condition constraint at time t, and δ is the confidence interval coefficient. A value of 1.96 corresponds to the 95% confidence interval, thus obtaining the initial dynamic baseline.

[0123] Subsequently, a real-time update mechanism was designed: a parameter iteration is triggered at a default interval, and real-time monitoring data (including deformation data and operating condition data) within a past period is selected as the new data D_new, which is concatenated with the historical data D to form D_update=[D,D_new]. An incremental learning strategy is used to update the posterior distribution P(θ|D_update,C_new), where C_new is the new operating condition factor vector. This avoids the computational delay caused by retraining with all data, thereby obtaining a real-time dynamic deformation baseline, which includes the baseline value and baseline fluctuation range of each monitoring point, providing an accurate benchmark for subsequent deformation anomaly judgment.

[0124] Furthermore, the step of introducing the dynamic deformation baseline features adapted to real-time operating conditions, the precise multi-source fusion features, and the spatiotemporal aggregated deformation features into a Kalman filter to construct an adaptive early warning threshold for operating conditions and complete monitoring and early warning includes the following steps:

[0125] S421. Using the dynamic deformation baseline as a reference, and combining the precise multi-source fusion features and the spatiotemporal aggregated deformation features, the deformation field is reconstructed through Kalman filtering.

[0126] First, feature fusion preprocessing is performed to weight and fuse the precise multi-source fusion feature Q and the spatiotemporal aggregated deformation feature W to generate observations. The fusion weights are dynamically allocated based on real-time operating conditions: when the mining intensity is >50t / h or within 30 minutes after blasting (highly dynamic deformation), the weight of the spatiotemporal aggregation deformation feature is set to 0.6 (highlighting the deformation trend information of spatiotemporal aggregation), and the weight of the precise multi-source fusion feature is set to 0.4 (ensuring basic accuracy after interference compensation); when the operating conditions are stable, both weights are set to 0.5 to balance basic accuracy and trend capture capability. The fusion formula is as follows: α is the weight of the precise multi-source fusion feature, with a value range of [0.4, 0.5]. After fusion, the feature distribution offset is eliminated by compression through a 1×1 convolutional layer and normalization by BatchNorm.

[0127] Subsequently, an improved Kalman filter model was constructed: the baseline values ​​B(x,t) of each monitoring point in the dynamic deformation baseline were used as the initial state values. Simultaneously, temporal trend information (such as strain rate of change and centroid offset rate) from the spatiotemporal aggregate deformation characteristics is introduced to calculate the deformation trend correction term. The formula is Δc is the temporal deformation rate, Δd is the spatial deformation gradient, and ω1 and ω2 are weighting coefficients, which are set to 0.6 and 0.4 respectively in the example.

[0128] Furthermore, optimize the design of the core equations:

[0129] Equations of state: The state transition matrix A employs an adaptive update mechanism. When the operating conditions are stable, A is set as a diagonal matrix (diagonal elements are 0.98 to preserve historical state stability). When the operating conditions change abruptly, the diagonal elements of A are adjusted to 0.92 (to enhance the ability to correct the current trend). The control matrix B is initially set as a 128×128 dimensional identity matrix and is iteratively optimized through backpropagation. Process noise... It follows a mean of 0 and a covariance matrix. Dynamically adjusted Gaussian distribution σi is determined by the variance of the prediction errors of nearly 50 sets of states. The larger the error, the larger σ_i, which improves the robustness of the model.

[0130] Observation equation: The observation matrix H is obtained by fitting the mapping relationship between the observed values ​​and the state values ​​using the least squares method, ensuring observation accuracy; observation noise It follows a mean of 0 and a covariance matrix. A fixed Gaussian distribution Based on sensor accuracy calibration, it is adapted to sub-millimeter level requirements.

[0131] Furthermore, the Kalman filter iterative process is optimized:

[0132] Prediction phase: Based on the state equation, the predicted state is calculated to satisfy: The prediction error covariance satisfies: ;

[0133] Update phase: Calculate Kalman gain Combined with observations Update status Synchronously update error covariance (where I is the identity matrix).

[0134] To avoid filter divergence, a residual verification mechanism is introduced: calculate the residual. ,like If the absolute value of the residual is greater than 3σ_v, where σ_v is the standard deviation of the residual, it is considered an observation anomaly. The mean of the first three effective residuals is used to replace the current residual to ensure iterative stability. The final output is the real-time deformation monitoring value of each monitoring point. Based on this value, Kriging interpolation is used to generate the global deformation field matrix of the goaf, thus completing the deformation field reconstruction.

[0135] S422. Based on the statistical characteristics of the monitoring data and the operating condition threshold, an adaptive early warning threshold is constructed. By combining the reconstruction results and the adaptive early warning threshold, deformation early warning with sub-millimeter accuracy is achieved.

[0136] In this embodiment, a sliding window mechanism is used to select nearly 100 sets of valid deformation monitoring values. The window sliding step size is set to 10 sets of data, that is, the statistical characteristics are updated once every 10 sets of data. If the residual test determines that there are abnormal monitoring values ​​in the window, cubic spline interpolation is used to replace them to ensure the purity of the statistical data.

[0137] Based on the data in this window, statistical characteristic parameters are calculated, including the mean, standard deviation, and skewness S of the data distribution. These parameters are used to determine whether the deformation data has a skewed distribution. For example, deformation after blasting is prone to positive skewness. If |S| > 0.5, the Box-Cox transformation is used to correct the data distribution and ensure the reliability of the statistical characteristics.

[0138] Subsequently, a working condition correction term Δg is introduced. Δg is calculated from the real-time working condition factor vector C output by the dynamic baseline, with the formula Δg = β1·C1 + β2·C2 + β3·C3 + β4·C4, where C1 is the mining volume percentage (norm), C2 is the pressure per unit area (norm), C3 is the relative strain value (norm), and C4 is the peak ground acceleration (peak acceleration) (norm); β1-β4 are weighting coefficients, determined based on the analytic hierarchy process, and are 0.3, 0.25, 0.25, and 0.2, respectively. The value range of Δg is strictly constrained to [0.1mm, 0.5mm]: when the mining volume percentage > 70% or the peak ground acceleration > 1.5m / s². 2 When the volume of mining is less than 30% and there is no blasting operation, Δg is taken as 0.3~0.5mm; when the volume of mining is less than 30% and there is no blasting operation, Δg is taken as 0.1~0.2mm.

[0139] The final adaptive warning threshold satisfies: Where 3σ is the statistical redundancy term and Δg is the operating condition risk redundancy term. Multi-level early warning thresholds are defined based on the core threshold:

[0140] Level 1 warning threshold (High risk threshold, corresponding to high risk of deformation and collapse);

[0141] Level II warning threshold (Medium risk threshold, corresponding to a clear trend of accelerated deformation).

[0142] Attention threshold Low-risk threshold, corresponding to stable deformation but requiring enhanced monitoring.

[0143] Furthermore, to ensure the validity of the threshold, a dual verification mechanism is established:

[0144] Historical data verification: Compare the matching degree between the threshold and the historical warning data of the past 3 months. The warning accuracy rate is required to be ≥95% (historical warning accuracy rate = number of correct warnings / total number of warnings). If it is not met, the weight coefficients of β1-β4 will be adjusted.

[0145] Real-time operating condition verification: When a sudden change in operating condition is detected (such as peak blasting vibration > 2 m / s), 2 This immediately triggers a temporary threshold correction, simultaneously increasing T1, T2, and T3 by 10%. The original threshold is restored 30 minutes after the mutation, thus avoiding missed warnings under mutation conditions.

[0146] In addition, a dynamic threshold update mechanism is designed, which updates the threshold every 5 minutes by default and synchronizes the sliding window data. When the deformation monitoring value is detected to be close to T3 for 5 consecutive groups, the update cycle is shortened to improve the threshold's response speed to deformation trends.

[0147] Furthermore, a comprehensive decision-making mechanism is adopted, which includes real-time judgment at a single monitoring point, collaborative verification at multiple monitoring points, and graded response to early warnings, to improve the accuracy and reliability of early warnings.

[0148] First, perform real-time judgment at each monitoring point and record the real-time deformation monitoring values ​​of each monitoring point. The warning level for a single monitoring point is obtained by comparing it one by one with the three threshold levels (T1, T2, T3): ① If ≥T1, marked as Level 1 warning status (high risk); ② If T2≤ <T1, marked as Level II warning status (medium risk); ③If T3≤ <T2, marked as a state of concern (low risk); ④ If <T3, marked as normal state.

[0149] To avoid misjudgments caused by sensor failure at a single monitoring point, a single monitoring point status stability verification is introduced: if a monitoring point is marked as an early warning state (Level 1 / Level 2), three consecutive sets of data are required to maintain this state before the early warning can be confirmed as valid; if only a single set of data triggers the early warning, it is determined to be a momentary interference, marked as a suspected early warning, and the data is retained and monitored continuously.

[0150] Subsequently, multi-monitoring point collaborative verification was performed: Considering the spatial transmission of deformation in the goaf (e.g., deformation in one area can spread to adjacent areas), the goaf was divided into 5-8 key areas according to spatial clustering results. The percentage r of the number of monitoring points in the early warning state in each area was counted (r = number of early warning points / total number of monitoring points in the area): ① If r > 50% in a certain area and contains at least one Level 1 early warning point, a Level 1 early warning for the area is triggered; ② If 30% < r ≤ 50% in a certain area and there are no Level 1 early warning points, a Level 2 early warning for the area is triggered; ③ If multiple adjacent areas trigger early warnings at the same time, it is determined to be a region-wide early warning, and the early warning response level is upgraded (e.g., a Level 2 early warning for the area is upgraded to a Level 1 early warning for the entire region).

[0151] Finally, the following tiered early warning response is implemented: ① Level 1 Early Warning Response (High Risk): Immediately trigger audible and visual alarms (installed at the mining face and monitoring center), and simultaneously push early warning information (including the warning area, real-time deformation value, and operating data) to the monitoring center, mining equipment control system, and management personnel's mobile APP via industrial Ethernet; forcibly initiate the emergency shutdown procedure for mining equipment (linked through the PLC control system), prohibiting personnel from entering the warning area; dispatch professional personnel with portable monitoring equipment to the site to investigate potential hazards such as support structure damage and rock wall cracks. The warning can only be lifted after the hazards are eliminated and the deformation monitoring value drops below T3. ② Level 2 Early Warning Response (Medium Risk): Trigger the audible and visual alarm at the monitoring center and push the warning information to management personnel; increase the monitoring frequency and strengthen the capture of time-series deformation trends; arrange personnel to patrol the warning area once and record the deformation changes. ③ Attention Status Response (Low Risk): Do not trigger audible and visual alarms, only mark it as "attention" in the monitoring center system; maintain the normal monitoring frequency and generate a deformation trend report. ④ Normal Status Response: Maintain the routine monitoring and baseline update mechanism, output a deformation field report, and simultaneously verify the validity of the dynamic baseline to ensure the stable operation of the monitoring system.

[0152] Please see Figure 2 In this embodiment, to efficiently execute the high-precision monitoring and early warning method for deformation fields in underground mine goafs provided by this invention, the present invention also provides a high-precision monitoring and early warning system for deformation fields in underground mine goafs, comprising: an input device, an output device, a processor, and a memory, wherein the input device, output device, processor, and memory are interconnected, and the memory contains program instructions for the steps of the high-precision monitoring and early warning method for deformation fields in underground mine goafs. The high-precision monitoring and early warning system for deformation fields in underground mine goafs of this invention has a compact structure and stable performance, and can stably execute the high-precision monitoring and early warning method for deformation fields in underground mine goafs of this invention, further improving the overall applicability and practical application capability of this invention.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the present invention.

Claims

1. A high-precision monitoring and early warning method for deformation field of underground mine goaf, characterized in that, Includes the following steps: Integrate features from multiple data sources to obtain preliminary cross-modal fusion features; Based on the environmental interference compensation features, the cross-modal preliminary fusion features, and the multi-source data features, the accurate multi-source fusion features after interference compensation are obtained through attention-weighted fusion and feature selection optimization. By combining the temporal deformation trend characteristics and spatial distribution characteristics of the goaf, spatiotemporal aggregate deformation characteristics are extracted; The dynamic deformation baseline features adapted to real-time operating conditions, the precise multi-source fusion features, and the spatiotemporal aggregated deformation features are introduced into Kalman filtering to construct an operating condition adaptive early warning threshold to complete monitoring and early warning. Extracting the environmental disturbance compensation features includes the following steps: Construct an SVR interference quantization model; Based on the interference factor data, the influence of interference on the monitoring data is quantified by the SVR interference quantification model, and interference compensation coefficients are generated to obtain environmental interference compensation characteristics that include the interference compensation coefficients of each monitoring data. The process of obtaining accurate multi-source fusion features after interference compensation based on environmental interference compensation features, cross-modal preliminary fusion features, and multi-source data features, through attention-weighted fusion and feature selection optimization, includes the following steps: Based on the environmental interference compensation features, the cross-modal preliminary fusion features, and the multi-source data features, an attention mechanism is introduced to calculate the weights of each feature; L1 regularization is used to filter out redundant features to achieve multi-feature fusion and obtain accurate multi-source fusion features after interference compensation. The process of extracting spatiotemporal aggregated deformation features by combining the temporal deformation trend characteristics and spatial distribution characteristics of the goaf includes the following steps: By combining the temporal and spatial characteristics of the temporal deformation trend and spatial distribution of the goaf, the intrinsic relationship between temporal and spatial characteristics can be captured. By introducing residual connections to avoid information loss during feature aggregation, deep aggregation of spatiotemporal features is achieved based on the intrinsic correlation, resulting in spatiotemporal aggregated deformation features that contain the intrinsic correlation between temporal trends and spatial distribution. Extracting the temporal deformation trend characteristics of the goaf includes the following steps: Introducing a spatiotemporal attention mechanism to highlight deformation information at key time points; The gating unit of the LSTM network is optimized, and based on the deformation information, long-term deformation trend features containing deformation information of key time periods are extracted through the optimized LSTM network. Extracting the spatial distribution characteristics of the goaf includes the following steps: Introduce a monitoring point reliability weight to optimize the Kriging interpolation algorithm; Based on the optimized Kriging interpolation algorithm, K-means++ spatial clustering is used to divide the key deformation regions and extract spatial distribution features. The process of introducing the dynamic deformation baseline features adapted to real-time operating conditions, the precise multi-source fusion features, and the spatiotemporal aggregated deformation features into a Kalman filter to construct an adaptive early warning threshold for operating conditions and complete monitoring and early warning includes the following steps: Based on the dynamic deformation baseline, and combined with the precise multi-source fusion features and the spatiotemporal aggregated deformation features, deformation field reconstruction is achieved through Kalman filtering. Based on the statistical characteristics of monitoring data and operating condition thresholds, an adaptive early warning threshold is constructed. By combining the reconstruction results and the adaptive early warning threshold, deformation early warning with sub-millimeter accuracy is achieved. The analysis of the dynamic deformation baseline characteristics adapted to the real-time operating conditions includes the following steps: Quantitative chemical condition factors; Using operating condition factors as constraints, an improved Bayesian inference algorithm is introduced to construct a dynamic deformation baseline, and the characteristics of the dynamic deformation baseline adapted to real-time operating conditions are analyzed.

2. The method according to claim 1, characterized in that, The process of integrating multi-source data features to obtain preliminary cross-modal fusion features includes the following steps: Extracting local features of a single modality; Based on the single-modal local features, a Transformer encoder is introduced to achieve global correlation fusion, thereby obtaining cross-modal preliminary fusion features that include local details of each modality and global correlation.

3. A high-precision monitoring and early warning system for deformation field of underground mine goaf, characterized in that, The high-precision monitoring and early warning system for deformation field in underground mine goaf includes: an input device, an output device, a processor, and a memory. The input device, output device, processor, and memory are interconnected. The memory includes program instructions, which are used to execute the high-precision monitoring and early warning method for deformation field in underground mine goaf as described in claim 1 or 2.

Images