Rockburst risk prediction and main control factor identification method based on SSA-attention-bilstm

Abstract

A rock burst risk prediction and master control factor identification method based on SSA-Attention-BiLSTM, which unifies geological data, mining area data and microseismic event level data into daily scale multi-source feature vectors, after data preprocessing and time alignment, constructs a time sequence sample set with continuous T-day features as input and T+1-day risk level as label; constructs BiLSTM to capture bidirectional time sequence dependence, introduces attention mechanism to focus on key time steps, and optimizes hyperparameters by sparrow search algorithm to obtain the optimal warning model of SSA-Attention-BiLSTM, which outputs four types of probabilities: no danger, weak danger, medium danger and strong danger; on high-risk samples, the attention weight is used to locate the key time window, the SHAP marginal contribution and feature ablation performance decay are weighted and fused, the comprehensive contribution degree of each factor is calculated and sorted, and the master control factor sequence containing factor category, key window, contribution degree and ablation influence is output. The method can realize accurate prediction of rock burst risk level and quantitative and accurate identification of master control factors.

Description

Technical Field

[0001] This invention belongs to the fields of mine dynamic disaster monitoring and early warning, deep learning time series prediction and interpretable machine learning technology, specifically involving a method for predicting rockburst risk and identifying the main controlling factors based on SSA-Attention-BiLSTM. Background Technology

[0002] Rockbursts are a typical dynamic disaster in deep coal mining, characterized by their suddenness, destructiveness, numerous contributing factors, strong coupling, and complex temporal evolution. With increasing mining depth, the stress environment of the coal and rock mass becomes increasingly complex, and the risk of rockbursts continues to rise, becoming a core bottleneck restricting safe coal mine production.

[0003] Rockbursts are the result of the coupled effects of multiple factors, including geological conditions, mining disturbances, and microseismic responses. Geological conditions determine the energy storage capacity and stress concentration basis of coal and rock masses; mining disturbances alter the stress field distribution around the working face and roadways; and microseismic responses directly reflect the evolution of coal and rock fracturing and the energy release process. Relying solely on a single type of indicator for early warning often only captures the surface symptoms of the disaster, failing to reveal its underlying mechanisms and leading to one-sided and insufficiently interpretable early warning results.

[0004] Existing research on rockbursts suffers from two significant technical limitations. First, the identification of controlling factors relies heavily on theoretical inference, empirical induction, and mathematical statistical methods. While these methods can compile a list of influencing factors and form a qualitative understanding, they cannot quantify the contribution of different factors to the actual risk assessment results. Second, although intelligent early warning models have widely incorporated machine learning and deep learning technologies, the research focus is mainly on improving the accuracy of risk level predictions. The interpretability of the models' output of high-risk results—specifically, identifying "which time periods are critical early warning windows" and "which factors are the core disaster-causing factors"—is severely lacking, making it difficult to guide precise on-site prevention and control operations.

[0005] Therefore, there is an urgent need to propose a method for identifying the main controlling factors of rockburst based on early warning model decision-making. On the basis of constructing a high-precision risk level early warning model by integrating geological, mining, and microseismic multi-source data, the contribution of various indicators to risk assessment is quantified through model interpretation and feature verification technology, so as to provide a scientific and comprehensive decision-making basis for rockburst prevention engineering design and precise on-site management. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a method for predicting rockburst risk and identifying the main controlling factors based on SSA-Attention-BiLSTM. This method can accurately predict the risk level of rockburst and can quantitatively and accurately identify the main controlling factors of rockburst by combining Attention weights, SHAP contribution, and feature ablation experiments, thus providing scientific and technical support for rockburst prevention engineering design and on-site treatment.

[0007] To achieve the above objectives, this invention provides a method for predicting rockburst risk and identifying controlling factors based on SSA-Attention-BiLSTM, comprising the following steps: Step 1: Multi-source data acquisition; around a specific working face or roadway area, acquire three types of data required for rockburst prediction: geological factors, mining factors, and microseismic monitoring characteristic indicators, and combine them into a daily unified multi-source feature vector; Step 2: Data preprocessing and time alignment; The three types of data generated are preprocessed and strictly aligned to time on a daily scale to obtain a unified geological, mining and microseismic multi-source characteristic sequence; Step 3: Constructing the sliding time window sample set; with continuous The multi-source feature sequence of day 1 is used as the model input, and the multi-source feature sequence of day 2 is used as the model input. The daily rockburst risk level is used as the output label. Samples are generated using a sliding window method and divided into training set, validation set and test set in chronological order. Step 4: Establishment of Attention-BiLSTM early warning model for rockburst risk level; Establish an Attention-BiLSTM model, use a bidirectional long short-term memory network to extract the long and short-term temporal dependencies of multi-source feature sequences, use the Attention mechanism to adaptively assign weights to different time steps to highlight key time step information, and finally output the probability of four risk levels through a fully connected layer and a Softmax layer. Step 5: Hyperparameter optimization based on the sparrow search algorithm; The sparrow search algorithm is used to automatically optimize the key hyperparameters of the Attention-BiLSTM model, and the validation set performance is used as the fitness target to obtain the optimal SSA-Attention-BiLSTM early warning model. Step 6: Risk level prediction; Input the multi-source time series samples to be predicted into the trained SSA-Attention-BiLSTM model, and output the risk level probability; Step 7: Identification of the main control factors based on model decision-making; Based on the risk discrimination results of the trained SSA-Attention-BiLSTM early warning model, combined with Attention weight positioning of key time windows, and by comprehensively applying SHAP feature contribution analysis and feature ablation experiments, the contribution degree and intensity of each indicator in geological factors, mining factors and microseismic monitoring feature indicators are quantified, and finally the sequence of main control factors of rockburst is identified.

[0008] As a preferred option, the multi-source data acquisition process in step 1 is as follows: S11: Acquire and compile geological and mining data; taking the target working face or roadway area as the research object, collect and organize daily records of geological and mining factors on a daily scale; among which, geological factor data includes burial depth, fault distance, coal seam thickness and hard roof thickness, impact energy index or elastic energy index; mining factor data includes working face advance speed, mining height, distance from working face, distance from goaf, and distance from coal pillar; S12: Process raw microseismic data and generate daily-scale characteristic indices; export event-level raw data from the mine microseismic monitoring system, including event occurrence time, source spatial coordinates, event energy, magnitude, and waveform information; perform daily aggregation statistics and calculate the daily frequency of microseismic events. Total daily energy of micro-seismic events Maximum single energy of micro-earthquakes b value Seismic focal concentration As a characteristic indicator for microseismic monitoring; S13: Concatenate multi-source feature vectors; Let the first... Geological factor data vector of the day , No. Data vector of mining factors of the day , No. Microseismic monitoring characteristic index vector According to the following formula, the first... Multi-source feature vectors of the day : ; In the formula, This represents the total number of input features. Indicates the first Heavenly The observed values ​​of each input indicator.

[0009] As an alternative, in S11, for geological factors that are relatively fixed in the short term, the most recent survey value is used in the daily record until the data is updated.

[0010] As a preferred option, in S12, the daily frequency of microseismic events is... Calculate using the following formula: ;in Indicates the first Number of microseismic events within a day; Total daily energy of micro-seismic events Calculate using the following formula: ;in Indicates the first Heavenly The energy of a microseismic event; Maximum single energy of micro-earthquakes Calculate using the following formula: ; b value Determine according to the following procedure: (Based on the...) A preset time window is selected centered on the sky. The magnitude or energy level of all microseismic events within this window is collected, and the magnitudes are statistically analyzed to be no less than each threshold. Cumulative number of events According to the Gutenberg-Richter relation The least squares method was used to fit the linear segment of magnitude-frequency, and the fitting slope was taken. The absolute value of b at that time ,in, Indicates the magnitude or energy level of a microseismic event. and For fitting parameters; or directly obtain the b-value output by the microseismic monitoring system for the current day as... ; focal concentration Calculate using the following formula: ; In the formula, For the first Heavenly The source coordinates of the event, The arithmetic mean of the coordinates of the epicenters of all events on that day; As a preferred embodiment, in step 2, the data preprocessing and time alignment process is as follows: S21: Handling missing and outlier values; For short-term missing data, use linear interpolation or previous value imputation; For outliers that clearly exceed physical meaning or statistical distribution, use threshold truncation or removal followed by interpolation for smoothing. S22: Indicator normalization processing; use Min-Max normalization or Z-score standardization to normalize all features; Min-Max normalization is shown in the following formula: ,in and Features The minimum and maximum values ​​in the entire sample. To prevent extremely small positive numbers with a denominator of zero; Z-score standardization is shown in the following formula: ,in and They represent the first The mean and standard deviation of each feature in the training samples; S23: Daily-scale time alignment; ensure that the timestamps of all data records are unified to the daily granularity; check the time indexes of the three types of data (geological, mining, and microseismic) one by one, fill in missing dates, remove duplicate records, and form a continuous and aligned normalized multi-source feature sequence as a multi-source time series sample for constructing a sliding time window sample.

[0011] As a preferred embodiment, the sliding time window sample set construction process in step 3 is as follows: S31: Determine the risk level label; the risk level is determined based on historical impact events, records of severe mining pressure manifestations, on-site early warning records, expert judgment results, or comprehensive evaluation conclusions. Risk level Including no danger, low danger, medium danger, and high danger; S32: Sliding window constructs input / output pairs; sets the time window length to... The sliding step size is 1 day; from continuous The normalized multi-source feature vectors of the day constitute the first Input matrix of samples , by the The output label is composed of the risk level of the ground pressure on the sky. ;in, It is the set of real numbers; S33: Divide the dataset according to time order; arrange all samples in chronological order, and then divide them into training set, validation set and test set according to time order.

[0012] As a preferred embodiment, in step 4, the process of outputting the probabilities of the four risk levels through the fully connected layer and the Softmax layer is as follows: S41: Bidirectional LSTM timing coding; This converts the samples... Input the data into the BiLSTM network step by step, and calculate the hidden states of the feedforward LSTM. Hidden states of backward LSTM splicing together the first Two-way hidden state of the step , where [;] denotes vector concatenation operation; S42: Attention weight allocation; the learnable attention layer calculates the attention score for each time step based on the hidden state at each time step. Attention weights are obtained after normalization. ,in , , These are learnable parameters; S43: Global Representation and Risk Prediction; Utilizing Attention Weights We obtain the global temporal representation by weighted summation of the hidden states. ; Represent the global time series Input a fully connected layer and a softmax layer, output the risk probability. ,in, , , , These represent the probabilities of no danger, low danger, moderate danger, and high danger, respectively; the final risk level is the category corresponding to the highest probability. .

[0013] As a preferred option, the hyperparameter optimization process based on the sparrow search algorithm in step 5 is as follows: S51: Hyperparameter encoding and search space definition; encoding the hyperparameters to be optimized into individual position vectors. ,in For learning rate, The number of hidden cells in the BiLSTM. The length of the time window. This refers to the Dropout ratio. The L2 regularization coefficient is... Set the batch size; simultaneously, set a reasonable search range for each parameter. ; S52: Define the fitness function and model training; cross-entropy loss is used for classification tasks. The optimization objective is to minimize the loss on the validation set or maximize the macro-average F1 score, and the fitness function is: In each evaluation, based on the current... Reconstruct the samples and divide them into training and validation sets to complete the training and evaluation of the model; S53: Iterative Optimization and Determination of the Optimal Model; The sparrow search algorithm iteratively updates the population through a mechanism of global exploration by discoverers, local following by joiners, and anti-predation behavior by scouts, seeking the hyperparameter combination that optimizes fitness. After the iteration ends, with The SSA-Attention-BiLSTM early warning model was trained using these hyperparameters.

[0014] As a preferred option, the risk level prediction process in step 6 is as follows: For future continuity The multi-source time series data to be predicted for 1 day were preprocessed and constructed into samples according to the same process. Inputting the SSA-Attention-BiLSTM early warning model yields four types of risk probabilities. And the final prediction level, including the probability of four types of risks. It includes four categories: no danger, low danger, medium danger, and high danger.

[0015] As a preferred option, the process of identifying the controlling factors based on model-driven decision-making in step 7 is as follows: S71: Attention-based key time window localization; extracting high-risk samples with true labels of strong and moderate risk from the test set or all samples, and assigning attention weights to these samples. The average weights are calculated over time steps, and the top-k time steps or consecutive high-weight intervals with the highest average weights are selected as the key time windows for the model to identify high risks. S72: Calculation of global and local feature contributions based on SHAP; within the key time window and global sample range, the SHAP method is applied to calculate the marginal contribution of each feature; The mean absolute SHAP value of each feature is ,in For the first The first sample The SHAP value of each feature; S73: Integration of Feature Ablation Experiments and Overall Contribution; Conduct feature ablation experiments, using the performance of the complete model as a benchmark, sequentially remove or zero out individual features to obtain the post-ablation performance. Calculate the ablation effect of this feature. ;like ,make ; Combine SHAP values ​​with ablation results to calculate the first The combined contribution of each factor ,in These are the weighting coefficients; S74: Ranking of key control factors and output of anti-collision suggestions; sort all factors by... The factors are sorted from largest to smallest to form a sequence of controlling factors. The identification results include factor name, category, corresponding key time window, SHAP contribution, ablation performance decline, and overall contribution. For factors with high contribution, targeted anti-shock auxiliary decision-making suggestions are proposed.

[0016] As a preferred option, anti-slip auxiliary decision-making recommendations include: When the contribution of distance from the fault or thickness of the hard roof is high, it suggests strengthening the regional anti-rockfall design; when the contribution of working face advance speed or distance from the working face to the coal pillar is high, it suggests optimizing the mining rhythm and focusing on controlling high-stress areas; when the contribution of daily total energy of microseismic events, daily maximum single energy of microseismic events, b-value, or source concentration is high, it suggests closely monitoring abnormal energy release and areas of source concentration.

[0017] To address the challenges of numerous, highly coupled, and complex temporal evolution factors contributing to rockbursts, traditional engineering identification methods struggle to accurately identify the controlling factors. Furthermore, existing rockburst early warning research often focuses on risk level prediction, failing to adequately explain the contributions of geological factors, mining factors, and microseismic monitoring indicators in risk assessment. This invention provides a rockburst risk prediction and controlling factor identification method based on SSA-Attention-BiLSTM. First, addressing the inconsistency in spatiotemporal scales between geological factors (relatively static), mining factors (dynamically changing), and microseismic monitoring (event-level high-frequency) during rockburst formation, this invention converts raw microseismic event data into daily-scale statistical features (frequency, total energy, maximum energy, b-value, source concentration), strictly aligning them daily with geological and mining indicators to construct continuous, standardized multi-source time-series samples. This preserves the dynamic evolution information of the microseismic response while achieving unified modeling of three types of heterogeneous data, providing a high-quality data foundation for subsequent deep learning. Secondly, preprocessing and time-aligning geological factors, mining factors, and microseismic monitoring indicators on a daily scale provides a clean and unified data foundation for the subsequent construction of multi-source time-series samples. Furthermore, a rockburst risk level early warning model is established by integrating geological factors, mining factors, and microseismic monitoring indicators. A bidirectional long short-term memory network is used to extract the long-short-term dependencies of multi-source indicators within continuous time windows, overcoming the shortcomings of traditional methods in capturing the temporal evolution of rockburst precursors. Simultaneously, an attention mechanism is introduced to adaptively assign weights to different time steps, highlighting the key time steps that contribute most to risk assessment and enhancing the model's interpretability. Subsequently, a sparrow search algorithm is used to automatically optimize key hyperparameters such as learning rate, number of hidden units, time window length, Dropout ratio, regularization coefficient, and batch size, avoiding the blindness of manual parameter tuning and significantly improving the model's generalization performance on the validation set. Moreover, it exhibits higher early warning accuracy and robustness compared to traditional models. Next, the trained early warning model is used to predict the risk level of rockbursts. It can accurately predict four risk levels: no risk, weak risk, moderate risk, and strong risk. This facilitates the implementation of corresponding emergency response measures for different levels, ensuring timely and accurate handling. Finally, based on the decision results of the early warning model, the Attention weight is used to locate the key time window for high-risk assessment, answering the question of when is important. The SHAP feature contribution (global and local) is used to quantify the marginal contribution of each geological, mining, and microseismic indicator to the risk level assessment, reflecting the magnitude of each factor's influence in a unified dimension. Feature ablation experiments are used to sequentially remove individual indicators, verifying the model's performance degradation and further confirming the indispensability of each indicator. The SHAP values ​​and ablation results are weighted and fused to form a comprehensive contribution, ultimately outputting the ranking and intensity of the controlling factors.Therefore, combining the SHAP method and feature ablation experiments can effectively quantify the contribution of geological, mining, and microseismic indicators in risk assessment, and generate ranking and verification results of the main control factors. This breaks through the limitation of existing early warning research, which only focuses on risk level classification, and realizes the transformation from early warning of rockburst risk levels to quantitative identification of main control factors and quantitative interpretation of factor contributions. It enables field engineers to clearly understand the basis for the model's high-risk judgment, providing a scientific basis for rockburst prevention engineering design, mining rhythm optimization, and targeted on-site management. This invention implicitly captures the nonlinear interactions between factors through an end-to-end deep learning model, and uses Attention, SHAP, and ablation experiments to invert complex model decisions into an understandable ranking of factor contributions. It effectively solves the technical problem of difficulty in quantifying and decomposing the coupling effect of geology, mining, and microseismicity, which is crucial for identifying the main control factors of rockburst, optimizing rockburst prevention measures, and improving on-site targeted management capabilities.

[0018] This method can accurately predict the risk level of rockburst. Furthermore, based on the early warning of rockburst risk level, it further quantifies the contribution of geological factors, mining factors, and microseismic monitoring characteristic indicators to the rockburst risk level discrimination results, clarifies the role of different time windows and different influencing factors in the evolution of rockburst risk, and achieves the unity of rockburst risk level prediction and quantitative identification of main controlling factors, providing an innovative technical path for intelligent and precise prevention and control of rockburst. Attached Figure Description

[0019] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram of the process for identifying the main control factors in decision-making based on the early warning model in this invention. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] This invention provides a method for predicting rockburst risk and identifying controlling factors based on SSA-Attention-BiLSTM, comprising the following steps: Step 1: Acquisition of multi-source data; Around a specific working face or roadway area, three types of data are acquired: geological factors, mining factors, and microseismic monitoring characteristic indicators required for rockburst prediction. These data are then combined into a unified multi-source feature vector for each day, providing complete input variables for subsequent time series modeling. In one implementation, a specific working face or tunnel area is taken as the research object, and geological factors, mining factors, and microseismic monitoring factors are organized on a daily scale. Specifically, the multi-source data acquisition process is as follows: S11: Acquire and compile geological and mining data; taking the target working face or roadway area as the research object, collect and organize daily records of geological and mining factors on a daily scale; among which, geological factor data includes burial depth, fault distance, coal seam thickness and hard roof thickness, impact energy index or elastic energy index; mining factor data includes working face advance speed, mining height, distance from working face, distance from goaf, and distance from coal pillar; For geological factors that are relatively fixed in the short term (such as burial depth and coal seam thickness), the most recent survey values ​​can be used in the daily records until the data is updated. S12: Process raw microseismic data and generate daily-scale characteristic indicators; since raw microseismic monitoring data is event-level data, while geological and mining factors are mostly described according to working faces or areas, it is necessary to convert the raw microseismic event-level data into daily-scale statistical characteristic indicators before constructing time-series samples; specifically, export event-level raw data from the mine microseismic monitoring system, including event occurrence time, source spatial coordinates, event energy, magnitude, and waveform information; perform daily aggregation statistics and calculate the daily frequency of microseismic events. Total daily energy of micro-seismic events Maximum single energy of micro-earthquakes b value Seismic focal concentration ; Daily frequency of microseismic events Calculate using the following formula: ;in Indicates the first Number of microseismic events within a day; Total daily energy of micro-seismic events Calculate using the following formula: ;in Indicates the first Heavenly The energy of a microseismic event is measured in joules (J). Maximum single energy of micro-earthquakes Calculate using the following formula: ; b value Determine according to the following procedure: (Based on the...) A preset time window is selected centered on the sky. The magnitude or energy level of all microseismic events within this window is collected, and the magnitudes are statistically analyzed to be no less than each threshold. Cumulative number of events According to the Gutenberg-Richter relation The least squares method was used to fit the linear segment of magnitude-frequency, and the fitting slope was taken. The absolute value of b at that time ,in, Indicates the magnitude or energy level of a microseismic event. and For fitting parameters; or directly obtain the b-value output by the microseismic monitoring system for the current day as... ; focal concentration (Reflecting the degree of spatial clustering of events, expressed as the root mean square difference of the daily epicenter coordinates; the smaller the value, the more concentrated the spatial distribution of events.) Calculated according to the following formula: ; In the formula, For the first Heavenly The source coordinates of the event, The arithmetic mean of the coordinates of the epicenters of all events on that day, in meters; S13: Concatenate multi-source feature vectors; combine the above 15 indicators by day; let the first... Geological factor data vector of the day , No. Data vector of mining factors of the day , No. Microseismic monitoring characteristic index vector According to the following formula, the first... Multi-source feature vectors of the day : ; In the formula, This represents the total number of input features. Indicates the first Heavenly The observed values ​​of each input indicator.

[0022] This technical solution unifies the daily-scale data describing geological and mining operations by working face or region with the event-level raw data captured in real time by the microseismic monitoring system into a standardized multi-source feature vector with daily granularity. This resolves the inherent gaps in time scale, data structure, and physical dimensions between multi-source heterogeneous data. In particular, by aggregating the raw microseismic event-level records into five statistical characteristic indicators—daily frequency, daily total energy, daily maximum single energy, b-value, and source concentration—it not only fully preserves the temporal evolution and spatial clustering information of microseismic activity but also achieves precise time alignment with slowly changing geological and mining factors. This provides a structurally unified, informationally complete, and scale-consistent basic data matrix for subsequent construction of sliding time window samples and training of time-series early warning models.

[0023] Step 2: Data preprocessing and time alignment; Data preprocessing (data cleaning, missing value handling, outlier removal and normalization) is performed on the three types of generated data, and time alignment is strictly performed according to the daily scale to obtain a unified geological, mining and microseismic multi-source characteristic sequence, providing a clean and unified data foundation for subsequent processing; As a preferred approach, the data preprocessing and time alignment process is as follows: S21: Handling missing and outlier values; For short-term missing data, use linear interpolation or previous value imputation; For outliers that significantly exceed the physical meaning or statistical distribution (such as three times the standard deviation), use threshold truncation or removal followed by interpolation for smoothing. S22: Indicator normalization processing; To eliminate dimensional differences, Min-Max normalization or Z-score standardization methods are used to normalize all features; Min-Max normalization is shown in the following formula: ,in and Features The minimum and maximum values ​​in the entire sample. To prevent extremely small positive numbers with a denominator of zero; Z-score standardization is shown in the following formula: ,in and They represent the first The mean and standard deviation of each feature in the training samples; S23: Daily-scale time alignment; ensure that the timestamps of all data records are unified to the daily granularity; check the time indexes of geological, mining, and microseismic data one by one, fill in missing dates, remove duplicate records, and form a continuous and aligned normalized multi-source feature sequence as a multi-source time series sample for constructing sliding time window samples. This technical solution integrates missing value repair, outlier smoothing, multi-index normalization, and strict diurnal time alignment into a coherent data cleaning pipeline, systematically eliminating inconsistencies in sampling quality, dimensional scale, and temporal granularity among geological, mining, and microseismic data from different sources. This process not only ensures the numerical stability and comparability of input features but also guarantees a strict one-to-one correspondence of all variables on the time axis, thus providing a clean, complete, and physically causally matched standardized multi-source feature sequence for subsequent sliding time window truncation and BiLSTM time series modeling.

[0024] Step 3: Constructing the sliding time window sample set; With continuous The multi-source feature sequence of day 1 is used as the model input, and the multi-source feature sequence of day 2 is used as the model input. The risk level of rockburst within a day or a preset prediction time interval is used as the output label. Samples are generated using a sliding window method, and the training set, validation set, and test set are strictly divided in chronological order. As a preferred option, the process of constructing the sliding time window sample set is as follows: S31: Determine the risk level label; the risk level is determined based on historical impact events, records of severe mining pressure manifestations, on-site early warning records, expert judgment results, or comprehensive evaluation conclusions. Risk level Including no danger, low danger, medium danger, and high danger; S32: Sliding window constructs input / output pairs; sets the time window length to... The sliding step size is 1 day; from continuous The normalized multi-source feature vectors of the day constitute the first Input matrix of samples , by the The output label is composed of the risk level of the ground pressure on the sky. ;in, It is the set of real numbers; S33: Divide the dataset according to time order; arrange all samples strictly in chronological order, and then divide them into training set, validation set and test set in chronological order; the training set accounts for 70% of the total number of samples and is used for model learning; the validation set accounts for 15% of the total number of samples and is used for hyperparameter tuning and model selection; the test set accounts for 15% of the total number of samples and is used for final performance evaluation. Any disruption of the time order across sets is strictly prohibited to avoid future information leakage.

[0025] In this technical solution, continuous The multi-source feature sequence of day 1 is used as input, followed by the next day's... Using the risk level as a label, supervised learning samples were constructed to conform to the physical logic of rockburst time-series early warning, ensuring the temporal causality between input and output. Simultaneously, a dataset construction method that strictly divides the training, validation, and test sets according to chronological order fundamentally eliminates the leakage of future information caused by random shuffling. This ensures that the data distribution faced by the model during the validation and testing phases realistically simulates actual early warning scenarios that infer the future from historical data, thereby guaranteeing the objectivity of model performance evaluation and the reliability of the model's generalization ability in real-world applications.

[0026] Step 4: Establishment of Attention-BiLSTM rockburst risk level early warning model; An Attention-BiLSTM model is established, which includes an input layer, a BiLSTM layer, an Attention layer, a fully connected layer, and a Softmax output layer. A Bidirectional Long Short-Term Memory (BiLSTM) network is used to extract the long-short-term temporal dependencies of multi-source feature sequences. The Attention mechanism is used to adaptively assign weights to different time steps, highlighting key time step information that contributes significantly to risk assessment. Finally, four risk level probabilities are output through the fully connected layer and the Softmax layer: no risk, weak risk, moderate risk, and strong risk. As a preferred approach, the process of outputting the probabilities of four risk levels through a fully connected layer and a Softmax layer is as follows: S41: Bidirectional LSTM timing coding; This converts the samples... Input the data into the BiLSTM network step by step, and calculate the hidden states of the feedforward LSTM. Hidden states of backward LSTM splicing together the first Two-way hidden state of the step Where [;] denotes vector concatenation; the hidden state sequence output by BiLSTM is ; S42: Attention weight allocation; the learnable attention layer calculates the attention score for each time step based on the hidden state at each time step. Attention weights are obtained after normalization. ,in , , For learnable parameters; attention weights Indicates the first The contribution weight of each time step to the risk assessment result directly reflects the importance the model attaches to each time step during the assessment, and can be used to identify the key time windows that the model focuses on in high-risk assessment.

[0027] S43: Global Representation and Risk Prediction; Utilizing Attention Weights We obtain the global temporal representation by weighted summation of the hidden states. ; Represent the global time series Input a fully connected layer and a softmax layer, output the risk probability. ,in, , , , These represent the probabilities of no danger, low danger, moderate danger, and high danger, respectively; the final risk level is the category corresponding to the highest probability. ; This technical solution utilizes a bidirectional LSTM to simultaneously extract multi-source time-series features from both forward and backward directions, fully capturing the bidirectional dependencies of rockburst precursor information along the time axis. This avoids the shortcomings of unidirectional models in insufficient utilization of long-term historical information. Furthermore, an attention mechanism is introduced, enabling the model to adaptively assign differentiated weights to different time steps, automatically focusing on the critical periods most discriminative for risk assessment. This improves early warning accuracy while also providing the model with interpretability—the attention weights directly reveal the temporal basis for high-risk judgments, providing decision-making clues for subsequent key time window positioning and identification of controlling factors.

[0028] Step 5: Hyperparameter optimization based on the sparrow search algorithm; The Sparrow Search Algorithm (SSA) is used to automatically optimize key hyperparameters of the Attention-BiLSTM model, such as learning rate, number of hidden units, and time window length. The optimal SSA-Attention-BiLSTM early warning model is obtained with validation set performance as the fitness target. As a preferred option, the hyperparameter optimization process based on the sparrow search algorithm is as follows: S51: Hyperparameter encoding and search space definition; encoding the hyperparameters to be optimized into individual position vectors. ,in For learning rate, The number of hidden cells in the BiLSTM. The length of the time window. This refers to the Dropout ratio. The L2 regularization coefficient is... Set the batch size; simultaneously, set a reasonable search range for each parameter. ; S52: Define the fitness function and model training; cross-entropy loss is used for classification tasks. The optimization objective is to minimize the loss on the validation set or maximize the macro-average F1 score, and the fitness function is: In each evaluation, based on the current... Reconstruct the samples and divide them into training and validation sets to complete the training and evaluation of the model; S53: Iterative Optimization and Determination of the Optimal Model; The sparrow search algorithm iteratively updates the population through a mechanism of global exploration by discoverers, local following by joiners, and anti-predation behavior by scouts, seeking the hyperparameter combination that optimizes fitness. After the iteration ends, with As the final hyperparameters, the SSA-Attention-BiLSTM early warning model is trained; In this technical solution, six key hyperparameters—learning rate, number of hidden units, time window length, Dropout ratio, regularization coefficient, and batch size—are uniformly encoded as optimization individuals in a sparrow search algorithm. Through a three-pronged collaborative search mechanism—global exploration by the discoverer, local following by the joiner, and anti-predation by the scout—the optimal combination is automatically and efficiently approximated in the continuous hyperparameter space, avoiding the blindness and time-consuming nature of manual trial and error. Simultaneously, the fitness function is based on the cross-entropy loss of the validation set or the macro-average F1 score, and samples are reconstructed according to the current time window length during each evaluation. This ensures the objectivity of hyperparameter evaluation and the synergistic matching between the optimal time window and model parameters, ultimately achieving optimal performance in parameter configuration for the SSA-Attention-BiLSTM early warning model.

[0029] Step 6: Risk level prediction; Input the multi-source time series samples to be predicted into the trained SSA-Attention-BiLSTM model, and output the probability of four risk levels: no risk, weak risk, moderate risk and strong risk. As a preferred option, the risk level prediction process is as follows: For future continuity The multi-source time series data to be predicted for 1 day were preprocessed and constructed into samples according to the same process. Inputting the SSA-Attention-BiLSTM early warning model yields four types of risk probabilities. And the final prediction level, including the probability of four types of risks. It includes four categories: no danger, low danger, medium danger, and high danger; As a preferred approach, to verify the effectiveness of the optimization and attention mechanisms, LSTM, BiLSTM, and Attention-BiLSTM were simultaneously built as comparative models under the same dataset, feature system, and time window construction method. Accuracy, precision, recall, F1 score, and area under the ROC curve were used for evaluation on the test set. Furthermore, quantitative comparisons were conducted with the comparative models to verify the improvement effect of SSA hyperparameter optimization and the attention mechanism on the early warning capability of rockburst risk. The early warning performance of each model for rockburst risk level can be quantitatively compared.

[0030] In this technical solution, the trained SSA-Attention-BiLSTM model is directly applied to future continuous... The risk simulation based on multi-source data can output in real time the probability and certainty levels of four categories: no danger, low danger, moderate danger, and high danger, providing quantitative and actionable early warning decision-making basis for on-site operations. Simultaneously, by systematically comparing LSTM, BiLSTM, and Attention-BiLSTM models on the same data benchmark, the system quantitatively verifies the dual benefits of hyperparameter optimization and the attention mechanism in early warning performance from multiple dimensions, including accuracy, macro / micro average F1 score, and AUC for each category. This ensures that the conclusions regarding model performance improvement are based on objective and reproducible statistical evidence, enhancing the scientific credibility of the early warning results. Step 7: Identification of the controlling factors for model-based decision-making; Based on the risk discrimination results of the trained SSA-Attention-BiLSTM early warning model, and combined with Attention weights to locate key time windows, within the key time windows and the global sample range, SHAP feature contribution analysis and feature ablation experiments are comprehensively used to quantify the contribution degree and intensity of each indicator among geological factors, mining factors and microseismic monitoring feature indicators, and finally identify the main controlling factor sequence and intensity of rockburst. As a preferred approach, the process of identifying the controlling factors in model-based decision-making is as follows: S71: Attention-based key time window localization; extracting high-risk samples with true labels of strong and moderate risk from the test set or all samples, and assigning attention weights to these samples. The average weights are calculated over time steps, and the top-k time steps or consecutive high-weight intervals with the highest average weights are selected as the key time windows for the model to identify high risks. S72: Calculation of global and local feature contributions based on SHAP; within the key time window and global sample range, the SHAP method is applied to calculate the marginal contribution of each feature; The mean absolute SHAP value of each feature is ,in For the first The first sample The SHAP value of each feature, which uniformly reflects the magnitude of the feature's influence on the model output; S73: Integration of Feature Ablation Experiments and Overall Contribution; Conduct feature ablation experiments, using the full model performance (e.g., F1 score) as a benchmark, sequentially remove or zero out individual features to obtain the post-ablation performance. Calculate the ablation effect of this feature. ;like ,make ; Combine SHAP values ​​with ablation results to calculate the first The combined contribution of each factor ,in These are weighting coefficients used to balance the two importance measures; S74: Ranking of key control factors and output of anti-collision suggestions; The larger the value, the higher the value. The greater the contribution of each factor to the assessment of rockburst risk, the more factors are considered. The factors are sorted from largest to smallest to form a sequence of controlling factors. The identification results include factor name, category, corresponding key time window, SHAP contribution, ablation performance decline, and overall contribution. For factors with high contribution, targeted anti-rockburst auxiliary decision-making suggestions can be proposed based on site conditions. Targeted anti-rockburst auxiliary decision-making suggestions include, for example, when the contribution of distance from the fault or thickness of the hard roof is high, suggesting strengthening regional anti-rockburst design; when the contribution of working face advance speed or distance from the working face to the coal pillar is high, suggesting optimizing mining rhythm and focusing on controlling high-stress areas; when the contribution of daily total energy of microseismic events, daily maximum single energy of microseismic events, b-value, or source concentration is high, suggesting close attention to abnormal energy release and source concentration areas.

[0031] This technical solution constructs a multi-stage interpretability analysis framework that integrates time-series focus, dual attribution, and quantitative fusion. First, it utilizes the attention weights of high-risk samples to statistically locate key time windows, accurately pinpointing critical periods of disaster precursors from the global timeline, thus solving the problem of diluted disaster-causing time nodes in traditional global analysis. Second, it introduces the SHAP method within the key time windows and the global sample scope, quantifying the marginal contribution of each feature from the model's internal decision-making mechanism. Simultaneously, it combines feature ablation experiments to quantify the actual impact of each feature from the perspective of the model's external predictive performance degradation. Finally, it merges the SHAP contribution and ablation impact into a comprehensive contribution for ranking through weighted coefficients, balancing the credibility of the model's internal attribution with the sensitivity to practical engineering applications. This avoids the biases that may arise from a single method, ensuring that the final output sequence of controlling factors and targeted mitigation suggestions possess both statistical interpretability and withstand practical testing through feature removal.

[0032] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for predicting rockburst risk and identifying controlling factors based on SSA-Attention-BiLSTM, characterized in that, Includes the following steps: Step 1: Multi-source data acquisition; around a specific working face or roadway area, acquire three types of data required for rockburst prediction: geological factors, mining factors, and microseismic monitoring characteristic indicators, and combine them into a daily unified multi-source feature vector; Step 2: Data preprocessing and time alignment; The three types of data generated are preprocessed and strictly aligned to time on a daily scale to obtain a unified geological, mining and microseismic multi-source characteristic sequence; Step 3: Constructing the sliding time window sample set; With continuous The multi-source feature sequence of day 1 is used as the model input, and the multi-source feature sequence of day 2 is used as the model input. The daily rockburst risk level is used as the output label. Samples are generated using a sliding window method and divided into training set, validation set and test set in chronological order. Step 4: Establishment of Attention-BiLSTM early warning model for rockburst risk level; Establish an Attention-BiLSTM model, use a bidirectional long short-term memory network to extract the long and short-term temporal dependencies of multi-source feature sequences, use the Attention mechanism to adaptively assign weights to different time steps to highlight key time step information, and finally output the probability of four risk levels through a fully connected layer and a Softmax layer. Step 5: Hyperparameter optimization based on the sparrow search algorithm; The sparrow search algorithm is used to automatically optimize the key hyperparameters of the Attention-BiLSTM model, and the validation set performance is used as the fitness target to obtain the optimal SSA-Attention-BiLSTM early warning model. Step 6: Risk Level Prediction; Input the multi-source time series samples to be predicted into the trained SSA-Attention-BiLSTM model, and output the risk level probability. Step 7: Identification of the main control factors based on model decision-making; Based on the risk discrimination results of the trained SSA-Attention-BiLSTM early warning model, combined with Attention weight positioning of key time windows, and by comprehensively applying SHAP feature contribution analysis and feature ablation experiments, the contribution degree and intensity of each indicator in geological factors, mining factors and microseismic monitoring feature indicators are quantified, and finally the sequence of main control factors of rockburst is identified.

2. The method for predicting rockburst risk and identifying controlling factors based on SSA-Attention-BiLSTM according to claim 1, characterized in that, In step 1, the process of acquiring multi-source data is as follows: S11: Acquire and compile geological and mining data; taking the target working face or roadway area as the research object, collect and organize daily records of geological and mining factors on a daily scale; among which, geological factor data includes burial depth, fault distance, coal seam thickness and hard roof thickness, impact energy index or elastic energy index; mining factor data includes working face advance speed, mining height, distance from working face, distance from goaf, and distance from coal pillar; S12: Process raw microseismic data and generate daily-scale characteristic indices; export event-level raw data from the mine microseismic monitoring system, including event occurrence time, source spatial coordinates, event energy, magnitude, and waveform information; perform daily aggregation statistics and calculate the daily frequency of microseismic events. Total daily energy of micro-seismic events Maximum single energy of micro-earthquakes b value focal concentration As a characteristic indicator for microseismic monitoring; S13: Concatenate multi-source feature vectors; Let the first... Geological factor data vector of the day , No. Data vector of mining factors of the day , No. Microseismic monitoring characteristic index vector According to the following formula, the first... Multi-source feature vectors of the day : ； In the formula, This represents the total number of input features. Indicates the first Heavenly The observed values ​​of each input indicator.

3. The method for predicting rockburst risk and identifying controlling factors based on SSA-Attention-BiLSTM according to claim 2, characterized in that, In S11, for geological factors that are relatively fixed in the short term, their most recent survey values ​​are used in the daily records until the data is updated.

4. The method for predicting rockburst risk and identifying controlling factors based on SSA-Attention-BiLSTM according to claim 2, characterized in that, In S12, the daily frequency of microseismic events Calculate using the following formula: ;in Indicates the first Number of microseismic events within a day; Total daily energy of micro-seismic events Calculate using the following formula: ;in Indicates the first Heavenly The energy of a microseismic event; Maximum single energy of micro-earthquakes Calculate using the following formula: ; b value Determine according to the following procedure: (Based on the...) A preset time window is selected centered on the sky. The magnitude or energy level of all microseismic events within this window is collected, and the magnitudes are statistically analyzed to be no less than each threshold. Cumulative number of events According to the Gutenberg-Richter relation The least squares method was used to fit the magnitude-frequency linear segment, and the fitting slope was taken. The absolute value of b at that time ,in, Indicates the magnitude or energy level of a microseismic event. and For fitting parameters; or directly obtain the b-value output by the microseismic monitoring system for the day as... ; focal concentration Calculate using the following formula: ； In the formula, For the first Heavenly The source coordinates of the event, It is the arithmetic mean of the coordinates of the epicenters of all events on that day.

5. The method for predicting rockburst risk and identifying controlling factors based on SSA-Attention-BiLSTM according to claim 1, characterized in that, In step 2, the data preprocessing and time alignment process is as follows: S21: Handling missing and outlier values; For short-term missing data, use linear interpolation or previous value imputation; For outliers that clearly exceed physical meaning or statistical distribution, use threshold truncation or removal followed by interpolation for smoothing. S22: Indicator normalization processing; use Min-Max normalization or Z-score standardization to normalize all features; Min-Max normalization is shown in the following formula: ,in and Features The minimum and maximum values ​​in the entire sample. To prevent extremely small positive numbers with a denominator of zero; Z-score standardization is shown in the following formula: ,in and They represent the first The mean and standard deviation of each feature in the training samples; S23: Daily-scale time alignment; ensure that the timestamps of all data records are unified to the daily granularity; check the time indexes of the three types of data (geological, mining, and microseismic) one by one, fill in missing dates, remove duplicate records, and form a continuous and aligned normalized multi-source feature sequence as a multi-source time series sample for constructing a sliding time window sample.

6. The method for predicting rockburst risk and identifying controlling factors based on SSA-Attention-BiLSTM according to claim 1, characterized in that, In step 3, the process of constructing the sliding time window sample set is as follows: S31: Determine the risk level label; the risk level is determined based on historical impact events, records of severe mining pressure manifestations, on-site early warning records, expert judgment results, or comprehensive evaluation conclusions. Risk level Including no danger, low danger, medium danger, and high danger; S32: Sliding window constructs input / output pairs; sets the time window length to... The sliding step size is 1 day; from continuous The normalized multi-source feature vectors of the day constitute the first Input matrix of samples , by the The output label is composed of the risk level of the ground pressure on the sky. ;in, It is the set of real numbers; S33: Divide the dataset according to time order; arrange all samples in chronological order, and then divide them into training set, validation set and test set according to time order.

7. The method for predicting rockburst risk and identifying controlling factors based on SSA-Attention-BiLSTM according to claim 1, characterized in that, In step 4, the process of outputting the probabilities of the four risk levels through the fully connected layer and the Softmax layer is as follows: S41: Bidirectional LSTM timing coding; This converts the samples... Input the data into the BiLSTM network step by step, and calculate the hidden states of the feedforward LSTM. Hidden states of backward LSTM splicing together the first Two-way hidden state of the step , where [;] denotes vector concatenation operation; S42: Attention weight allocation; the learnable attention layer calculates the attention score for each time step based on the hidden state at each time step. Attention weights are obtained after normalization. ,in , , These are learnable parameters; S43: Global Representation and Risk Prediction; Utilizing Attention Weights We obtain the global temporal representation by weighted summation of the hidden states. ; Global temporal representation Input a fully connected layer and a softmax layer, output the risk probability. ,in, , , , These represent the probabilities of no danger, low danger, moderate danger, and high danger, respectively; the final risk level is the category corresponding to the highest probability. .

8. The method for predicting rockburst risk and identifying controlling factors based on SSA-Attention-BiLSTM according to claim 1, characterized in that, In step 5, the hyperparameter optimization process based on the sparrow search algorithm is as follows: S51: Hyperparameter encoding and search space definition; encoding the hyperparameters to be optimized into individual position vectors. ,in For learning rate, The number of hidden cells in the BiLSTM. The length of the time window. This refers to the Dropout ratio. The L2 regularization coefficient is... Set the batch size; simultaneously, set a reasonable search range for each parameter. ; S52: Define the fitness function and model training; cross-entropy loss is used for classification tasks. The optimization objective is to minimize the loss on the validation set or maximize the macro-average F1 score, and the fitness function is: ; In each evaluation, based on the current Reconstruct the samples and divide them into training and validation sets to complete the training and evaluation of the model; S53: Iterative Optimization and Determination of the Optimal Model; The sparrow search algorithm iteratively updates the population through a mechanism of global exploration by discoverers, local following by joiners, and anti-predation behavior by scouts, seeking the hyperparameter combination that optimizes fitness. After the iteration ends, with The SSA-Attention-BiLSTM early warning model was trained using these hyperparameters.

9. The method for predicting rockburst risk and identifying controlling factors based on SSA-Attention-BiLSTM according to claim 1, characterized in that, In step 6, the risk level prediction process is as follows: For future continuity The multi-source time series data to be predicted for 1 day were preprocessed and constructed into samples according to the same process. Inputting the SSA-Attention-BiLSTM early warning model yields four types of risk probabilities. And the final prediction level, including the probability of four types of risks. It includes four categories: no danger, low danger, medium danger, and high danger.

10. The method for predicting rockburst risk and identifying controlling factors based on SSA-Attention-BiLSTM according to claim 1, characterized in that, In step 7, the process of identifying the controlling factors based on model-driven decision-making is as follows: S71: Attention-based key time window localization; extracting high-risk samples with true labels of strong and moderate risk from the test set or all samples, and assigning attention weights to these samples. The average weights are calculated over time steps, and the top-k time steps or consecutive high-weight intervals with the highest average weights are selected as the key time windows for the model to identify high risks. S72: Calculation of global and local feature contributions based on SHAP; within the key time window and global sample range, the SHAP method is applied to calculate the marginal contribution of each feature; The mean absolute SHAP value of each feature is ,in For the first The first sample The SHAP value of each feature; S73: Integration of Feature Ablation Experiments and Overall Contribution; Conduct feature ablation experiments, using the performance of the complete model as a benchmark, sequentially remove or zero out individual features to obtain the post-ablation performance. Calculate the ablation effect of this feature. ;like ,make ; By fusing the SHAP value and ablation results, the first... The combined contribution of each factor ,in These are the weighting coefficients; S74: Ranking of key control factors and output of anti-impact suggestions; sort all factors by... The factors are sorted from largest to smallest to form a sequence of controlling factors. The identification results include factor name, category, corresponding key time window, SHAP contribution, ablation performance decline, and overall contribution. For factors with high contribution, targeted anti-shock auxiliary decision-making suggestions are proposed.

Images