Rock burst risk grading prediction method based on improved LSTM neural network

Abstract

The invention provides a rockburst risk grading prediction method based on an improved LSTM neural network. The rockburst risk grading prediction method comprises the following steps: S1, collecting rockburst influence factor data; s2, preprocessing the rock burst influence factor data; s3, constructing an improved LSTM, residual connection and SVR collaborative fusion neural network model; s4, performing model training and hyper-parameter optimization; s5, performing double-label coding on the rockburst risk level; s6, performing model performance evaluation and optimization; and S7, performing rockburst risk grading prediction application. According to the method, accurate, efficient and dynamic grading prediction of the rockburst risk can be achieved, the method is suitable for rockburst risk grading prediction of deep underground engineering such as tunnels, mine pits and underground workshops, accurate technical support can be provided for construction safety decision making and protective measure making, and casualties and property losses caused by rockburst disasters are effectively reduced.

Description

Technical Field

[0001] This invention relates to the fields of geotechnical engineering safety prediction and artificial intelligence technology, and in particular to a rockburst risk classification prediction method based on an improved LSTM neural network. Background Technology

[0002] Rockburst is a sudden disaster caused by the release of ground stress, abrupt changes in rock mechanical properties, and the superposition of construction disturbances during deep underground engineering construction. Accurate prediction of its risk level is crucial to ensuring the safety of construction personnel and preventing project damage. Existing rockburst risk prediction technologies mainly have the following shortcomings: 1. Traditional prediction methods, such as empirical formulas and fuzzy comprehensive evaluation, rely on manually setting weights, making it difficult to quantify the temporal dynamic correlation of rockburst influencing factors, such as the cumulative effect of construction disturbance over time. The predictions are highly subjective and have poor adaptability. 2. Single AI models have inherent defects: Pure LSTM models are prone to gradient vanishing during deep network training, resulting in the loss of early temporal features and limited ability to capture long-term dependencies; standalone SVR models can only achieve regression fitting of static features and cannot explore the dynamic changes of influencing factors; residual connections are only used as auxiliary structures to alleviate gradient problems and have not formed a collaborative optimization mechanism with other models. 3. Existing fusion models mostly use one-way feature transfer, such as LSTM output directly inputting into SVR, lacking two-way interactive feedback, resulting in inaccurate feature correction, especially weak ability to distinguish boundary risk levels such as weak-to-medium rockburst and medium-to-strong rockburst; 4. The model has insufficient generalization ability. When faced with different geological conditions such as hard rock / soft rock, high / low ground stress and construction scenarios, it is prone to misjudgment due to differences in feature distribution, and lacks a closed-loop mechanism for continuous iterative optimization. Summary of the Invention

[0003] This invention provides a rockburst risk classification prediction method based on an improved LSTM neural network, which solves the technical pain points of existing rockburst risk classification prediction methods, such as insufficient capture of temporal features, poor model synergy, weak boundary sample discrimination ability, and limited generalization ability.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: The rockburst risk classification prediction method based on the improved LSTM neural network includes the following steps: S1: Collect data on rockburst influencing factors to form a raw dataset of rockburst influencing factors; S2: Clean the original dataset of rockburst influencing factors, construct time-series samples, and standardize the data to obtain a standardized dataset. Divide the standardized dataset into training set, validation set, and test set. S3: Construct a collaborative fusion neural network model; S4: Based on the training set and validation set, the collaborative fusion neural network model is trained using the total loss function, and the optimal model is determined through hyperparameter optimization and early stopping strategy; S5: Perform dual-label encoding on the rockburst risk level to obtain discrete labels adapted to the improved LSTM classification and continuous labels adapted to SVR regression. Associate the discrete and continuous labels with the training set, validation set, and test set one by one. S6: Based on the test set, discrete labels and continuous labels, evaluate the classification performance and regression performance of the optimal model, and obtain the final collaborative fusion prediction model by judging the threshold. S7: Collect data on rockburst influencing factors for the new project, preprocess the data according to the standard in step S2, input the data into the final collaborative fusion prediction model, and output the rockburst risk classification prediction results for the new project.

[0005] In this specification, the S3 collaborative fusion neural network model sequentially includes an embedding layer, an improved LSTM layer, a residual connection algorithm module, an SVR module, an attention layer, a fully connected layer, and an output layer. The embedding layer maps the features of the standardized dataset to a high-dimensional feature space, outputting high-dimensional features. The improved LSTM layer is a two-layer stacked structure used to extract the temporal dependency features of rockburst influencing factors. The residual connection algorithm module is built between the two LSTM layers, simultaneously receiving the temporal features output from the first LSTM layer and the high-dimensional features output from the embedding layer, outputting corrected features. The SVR module receives the corrected features and outputs continuously corrected values ​​through regression operations. The attention layer receives the deep temporal features output from the second LSTM layer and the continuously corrected values ​​output from the SVR module, fuses and calculates attention weights, and weights and optimizes the deep temporal features to obtain fused features. The fully connected layer performs dimensionality reduction mapping on the fused features, and the output layer outputs the probability distribution of rockburst risk levels through the Softmax function.

[0006] In this specification, the rockburst risk classification prediction method based on the improved LSTM neural network also includes S8: tracking the actual rockburst situation of the new project to form an actual rockburst dataset, comparing the prediction results with the actual rockburst data, and achieving iterative optimization of the model through sample expansion, parameter fine-tuning or architecture adaptation.

[0007] In this specification, the specific rules for dual-label encoding in S5 are as follows: Rockburst risk levels are divided into four levels: Level I (no rockburst), Level II (weak rockburst), Level III (moderate rockburst), and Level IV (strong rockburst); discrete labels use a combination of integer encoding and one-hot encoding, with Level I corresponding to 0 and [1,0,0,0], Level II corresponding to 1 and [0,1,0,0], Level III corresponding to 2 and [0,0,1,0], and Level IV corresponding to 3 and [0,0,0,1]; continuous labels linearly map discrete labels to continuous values ​​in the interval [0,1], with Level I corresponding to 0, Level II to 0.33, Level III to 0.66, and Level IV to 1.0; the label for each time series sample is based on the actual rockburst level corresponding to the last data point of that sample.

[0008] In this specification, the bidirectional interaction mechanism between the improved LSTM layer and the residual connection algorithm module in S3 is as follows: The residual connection algorithm module fuses and corrects the temporal features output by the first layer of the improved LSTM with the high-dimensional features output by the embedding layer to obtain the corrected features, which are then input into the second layer of the improved LSTM, providing a low-noise feature foundation for deep temporal feature extraction; during model training, the cross-entropy loss of the improved LSTM is backpropagated to the residual connection algorithm module to optimize its weight parameters, so that the feature correction is adapted to the temporal feature extraction requirements of the improved LSTM.

[0009] In this specification, the SVR module in S3 forms a closed-loop interaction with the improved LSTM layer and the residual connection algorithm module. Specifically: the SVR module receives the corrected features output by the residual connection algorithm module and outputs continuous corrected values ​​to the attention layer, providing a quantitative basis for feature weight optimization; the regression loss of SVR and the cross-entropy loss of the improved LSTM work synergistically to update the weight parameters of the residual connection algorithm module in reverse; simultaneously, the fused features optimized by the attention layer are fed back to the SVR module, adjusting its kernel function parameters to adapt to the feature distribution; the closed-loop interaction continues until the total model loss decreases by less than 1% for 10 consecutive training epochs. The iteration will end when the preset maximum number of iterations is reached.

[0010] In this specification, hyperparameter optimization in S4 adopts a grid search method. The hyperparameters to be optimized include the number of LSTM units, the dimension of fully connected layers, the learning rate, the number of iterations, the batch size, the LSTM regularization coefficient, the penalty coefficient of SVR, the kernel function parameters, and the insensitive loss coefficient. The average of the validation set accuracy and the F1 score are used as optimization targets. The early stopping strategy monitors the total loss or performance indicators of the validation set. If the total loss does not decrease or the performance does not improve for 10 consecutive training cycles, training is stopped and the current best model is saved.

[0011] In this specification, the classification performance evaluation metrics in S6 include accuracy, precision, recall, and F1 score, while the regression performance evaluation metrics include mean squared error and coefficient of determination. The thresholds for meeting the criteria are: accuracy ≥ 85%, F1 score ≥ 80%, coefficient of determination ≥ 0.8, and total loss ≤ 0.2. All thresholds must be met simultaneously to be determined as the final collaborative fusion prediction model. If the criteria are not met, the model should be retrained and evaluated after adjusting the model parameters, supplementing with time-series samples of the corresponding risk level, or optimizing the model architecture.

[0012] In this manual, the rockburst risk classification prediction results in S7 include core risk levels and auxiliary information. The core risk level is the level corresponding to the maximum value in the probability distribution. The auxiliary information includes the probability value of each risk level and the continuous correction value of SVR. If the maximum value in the probability distribution is <0.7, it is marked as a boundary risk sample, and the tendency level and corresponding enhanced monitoring recommendations are clearly stated.

[0013] In this specification, the data on rockburst influencing factors in S1 include: geostress parameters, rock mechanics parameters, engineering geological condition parameters, and construction disturbance parameters.

[0014] In summary, the present invention has at least the following beneficial effects: 1. More accurate capture of temporal features: Improved bidirectional interaction between LSTM and residual connections effectively alleviates gradient vanishing, fully preserves the correlation between early key features and later dynamic features, and better fits the dynamic formation mechanism of rockburst; 2. More reasonable feature correction and weight allocation: The continuous quantization correction value of SVR is fed back to the LSTM attention layer to dynamically adjust the feature weights, strengthen the contribution of factors that play a dominant role in rockburst risk, and significantly improve the ability to distinguish boundary risk levels. 3. Significantly improved prediction reliability and generalization ability: The collaborative optimization and full life cycle iteration mechanism of the three types of models enable the models to adapt to the differences in feature distribution under different geological conditions and construction scenarios, reduce the risk of misjudgment under extreme working conditions, and can be continuously optimized as engineering data accumulates; 4. Enhanced engineering practicality: The standardized data collection-preprocessing-prediction-verification process adapts to the real-time monitoring needs of engineering projects. The prediction results not only output the risk level but also provide quantitative probability and correction information, offering richer reference for construction safety decisions. 5. Enhanced intelligence: No manual intervention is required for feature weight setting, enabling end-to-end prediction from data input to risk level output, reducing reliance on the experience of professionals, and promoting the transformation of rockburst risk prevention and control from passive response to proactive early warning. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the rockburst risk classification prediction method based on an improved LSTM neural network involved in this invention.

[0016] Figure 2 This is a schematic diagram of the data acquisition and preprocessing process for rockburst influencing factors involved in this invention.

[0017] Figure 3 This is a schematic diagram illustrating the process of constructing the collaborative fusion neural network model involved in this invention.

[0018] Figure 4 This is a schematic diagram illustrating the final model prediction application and iteration process involved in this invention. Detailed Implementation

[0019] like Figure 1 As shown, this embodiment provides a rockburst risk classification prediction method based on an improved LSTM neural network, including: S1: Data Collection of Factors Affecting Rockburst The formation of rockburst risk is influenced by multiple factors, including in-situ stress, rock mechanical properties, engineering geological conditions, and construction disturbances. Therefore, it is necessary to collect complete and reliable raw data using standardized methods. The data collection and preprocessing process for rockburst influencing factors is as follows: Figure 2 As shown, the details are as follows: 1. Ground stress parameters: Field measurements were conducted using the hydraulic fracturing method or the stress relief method. A monitoring point was set up every 50 meters, and continuous monitoring was carried out for 72 hours. The maximum principal stress, minimum principal stress, and the difference between the two were recorded during the monitoring period, and the data accuracy was retained to 0.1 MPa. 2. Rock mechanical parameters: Laboratory tests were conducted according to the engineering rock test method standard. Representative rock samples from the engineering site were selected, with 3 parallel samples in each group. The average value of each parameter was obtained through uniaxial compressive strength test, elastic modulus test, and Poisson's ratio determination. The test error was controlled within 5%. 3. Engineering geological condition parameters: The rock integrity coefficient is calculated based on the acoustic wave test results; the joint density is calculated by measuring the number of joints within a 10m×10m area on-site and converting it to joints / meter. The degree of groundwater development is determined on-site according to four levels: none, weak, medium, and strong, which are quantified as 0, 0.3, 0.6, and 1.0, respectively.

[0020] 4. Construction disturbance parameters: Recorded in real time through the construction log, including: excavation speed, in meters / day; support method, shotcrete and anchor support quantified as 1.0, steel frame support quantified as 2.0, and combined support quantified as 3.0; blasting intensity, calculated as the product of explosive dosage (kg / blast) and blasting frequency (blasts / day), in kg / day.

[0021] This step outputs the original dataset of rockburst influencing factors, with the data format being (total number of records, 12), where 12 corresponds to 12 influencing factors: maximum principal stress, minimum principal stress, stress difference, uniaxial compressive strength, elastic modulus, Poisson's ratio, rock integrity coefficient, joint density, quantitative value of groundwater development, excavation speed, quantitative value of support method, and blasting intensity.

[0022] S2: Data Preprocessing of Rockburst Influencing Factors To adapt to the improved time-series input requirements of LSTM and the regression fitting needs of SVR, the original dataset needs to be systematically processed to ensure data quality and format consistency. Specific implementation details are as follows: 1. Data Cleaning: For missing values ​​of continuous parameters such as maximum principal stress and uniaxial compressive strength, the mean of the parameter in the entire dataset is used for filling; for missing values ​​of discrete parameters such as the quantified value of support method, the median is used for filling. Outliers in normally distributed data such as elastic modulus and Poisson's ratio are detected using the 3σ principle, and outliers in non-normally distributed data such as joint density and blasting strength are detected using the box plot method. After on-site verification, outliers are corrected using the moving average method of 5 adjacent data sets. If the verified data is invalid, it is directly removed. 2. Construction of time series samples: Since the improvement of LSTM needs to capture time series dependencies, the cleaned data is divided into time series samples according to the time order. Each sample consists of 10 consecutive sets of data. For example, in S3, the time series length T=10 and the sample format is (time series length, 12), ensuring that each sample contains enough dynamic change information. 3. Data Standardization: Z-score standardization is used to eliminate dimensional differences. The standardization formula is as follows: , where x is the cleaned data, μ is the mean of the feature, σ is the standard deviation, and after standardization, the mean of all features is 0 and the variance is 1, to avoid a feature from dominating model training due to an excessively large numerical range; 4. Dataset partitioning: The training set, validation set, and test set are partitioned using stratified sampling in a ratio of 7:2:1. During sampling, it is ensured that the distribution of rockburst risk levels in each dataset is completely consistent with that in the original dataset.

[0023] This step outputs a standardized dataset with a uniform format: the training set ( ,10,12); validation set ( ,10,12); test set ( (10, 12); where , , These represent the number of samples in each dataset, and will be used for model training in S4, hyperparameter optimization, and performance evaluation in S6, respectively.

[0024] S3: Improved LSTM + Residual Connection + SVR Collaborative Fusion Neural Network Model Construction 3.1 Overall Model Architecture The process of building a collaborative fusion neural network model is as follows: Figure 3 As shown, the overall model architecture is divided into seven layers according to the data flow. The functions of each layer are closely related to the algorithm: the input layer receives standardized data and passes it to the embedding layer; the embedding layer completes the mapping of low-dimensional features to high-dimensional space, providing a unified feature input for subsequent algorithms; the improved LSTM layer extracts temporal features through a two-layer stacked structure and connects with the residual to achieve bidirectional interaction between features and gradients; the SVR layer receives the features after residual correction, performs regression calculation, and outputs the corrected value to feed back to the attention layer; the attention mechanism layer integrates the SVR correction value to dynamically adjust the feature weights; the fully connected layer maps the fused features to a low-dimensional space; and the output layer outputs the probability distribution of rockburst risk level through the Softmax function.

[0025] 3.2 Input Layer and Feature Initialization The core function of the input layer is to receive standardized rockburst influencing factor data output by S2. The input data is defined in vector form as shown in formula (1): ;in Representing the Standardized values ​​of the factors influencing rockburst. The value range is 1 to , This represents the total number of factors influencing rockburst, corresponding to the 12 factors identified in S1; superscript. This is a matrix transpose operator used to convert row vectors into column vectors, adapting to the dimensionality requirements of subsequent levels of computation.

[0026] The standardized data from the input layer needs to be further mapped to a high-dimensional feature space; therefore, an embedding layer is introduced for feature transformation. The embedding layer achieves the mapping from low-dimensional to high-dimensional through fully connected operations, and the mapping formula is Equation (2): ;in The embedded high-dimensional feature vector has a dimension of . , Setting it to 128 allows for control of computational complexity while maintaining feature expressiveness. Here is the weight matrix of the LSTM embedding layer, with dimension 1. ; The bias vector for the LSTM embedding layer has a dimension of . This is used to adjust the feature offset after mapping, so as to avoid the feature distribution being too concentrated.

[0027] 3.3 Improved LSTM layer construction 3.3.1 Basic LSTM Unit State Update; The improved LSTM layer adopts a two-layer stacked structure, with 128 LSTM units in each layer. Through a gating mechanism—forget gate, input gate, and output gate—selective extraction and memorization of temporal features are achieved. The working process of a single LSTM unit is divided into state update and output calculation in sequence, with the formulas for each stage as follows: The forget gate is responsible for deciding whether to retain the cell state information from the previous moment. In rockburst prediction, some early influencing factors, such as initial ground stress, may continue to have an impact as construction progresses. Their weights need to be adjusted through the forget gate. Formula (3) is as follows: ;in This is the time step, with a value ranging from 1 to... Timing length Set to 10; For the first The LSTM forget gate output vector at time step 1 has a dimension of 1. ; Here is the LSTM forget gate weight matrix, with dimension 1. This is used to fuse the hidden state from the previous time step and the currently embedded features; For the first The LSTM hidden state vector at time step 1 has a dimension of 1. ; This is the LSTM forget gate bias vector; The Sigmoid activation function is expressed as follows: The output value ranges from 0 to 1 and is used to simulate the probability of forgetting or retaining.

[0028] The input gate is responsible for updating the cell state at the current moment and filtering out information in the currently embedded features that has a significant impact on rockburst risk. Formula (4) is as follows: ;in For the first The LSTM input gate output vector at time t has a dimension of . ; Here is the LSTM input gate weight matrix, with dimension 1. ; This is the input gate bias vector for the LSTM.

[0029] Candidate cell states are used to store new information at the current moment. The feature values ​​are mapped to the range of -1 to 1 through the hyperbolic tangent activation function to enhance the nonlinear expression of the features. Equation (5) is as follows: ;in For the first The candidate cell state vector at time step has a dimension of ; Here is the LSTM cell state weight matrix, with dimension 1. ; This is the LSTM cell state bias vector; The hyperbolic tangent activation function is expressed as follows: .

[0030] The current cell state is updated through the synergistic effect of the forget gate and the input gate, which retains key historical information while incorporating new current information. Formula (6) is as follows: ;in For the first The cell state vector at time step has a dimension of ; For the first Cell state vector at time step; The Hadamard product is used to perform element-wise operations on corresponding vectors, ensuring fine-grained state updates.

[0031] The output gate controls the output of the cell state, filtering out information that contributes to the hidden state at the current moment. Formula (7) is as follows: ;in For the first The LSTM output gate output vector at time t, with dimension t. ; The output gate weight matrix of the LSTM has a dimension of . ; This is the LSTM output gate bias vector.

[0032] No. The hidden state of the LSTM at time step 1 is the Hadamard product of the hyperbolic tangent of the output gate and the cell state. It is the core temporal feature extracted by the LSTM layer, and formula (8) is as follows: ;in For the first The LSTM hidden state vector at time step 1 has a dimension of 1. This will be used as the input data for the residual connection.

[0033] 3.3.2 Bidirectional interaction between residual connections and LSTM: Residual connections build direct paths, allowing the features of the embedding layer to be directly superimposed onto the hidden state of LSTM. This not only alleviates gradient vanishing but also preserves the original feature information, achieving bidirectional interaction with LSTM.

[0034] Forward computation of residual connections using LSTM hidden states and embedding layer features The input is the hidden state after residual correction, and the output is the hidden state. Formula (9) is: ;in The hidden state vector of the LSTM after residual correction has a dimension of . ; The residual connection weight matrix has dimensions of . This is used to adjust the contribution of the embedding layer features; Let be the residual connection bias vector, with dimension . .

[0035] The feedback effect of residual connections on LSTM is reflected in the input update of the second-layer LSTM: the residuals are corrected... As the input to the second LSTM layer, it replaces the traditional direct input of the hidden state of the previous layer, enabling the second LSTM layer to further extract deep temporal information based on the corrected high-quality features. The cell state and hidden state update formulas of the second LSTM layer are formulas (10) to (12), which are as follows: ; ; ;in The hidden state vector of the second LSTM layer has a dimension of ; The cell state weight matrix of the second LSTM layer has a dimension of . ; This is the cell state bias vector for the second layer of the LSTM. , , These are the forget gate, input gate, and output gate output vectors of the second-layer LSTM, respectively. Through this input replacement, the residual connection feeds the corrected features back to the deeper layers of the LSTM, ensuring that key information is not lost during the transmission of temporal features. It also alleviates the gradient vanishing problem. When the model backpropagates, the gradient can be directly transmitted to the shallow LSTM through the residual path, improving training stability.

[0036] The feedback from LSTM to residual connections is reflected in the parameter update process. During model training, the cross-entropy loss of the LSTM is backpropagated to the residual connections, and the weight matrix of the residual connections is optimized through gradient descent. This makes the residual correction more aligned with the feature extraction requirements of LSTM. The weight update formula (13) is: ;in The learning rate; The cross-entropy loss of LSTM; The partial derivative of the LSTM loss with respect to the residual connection weights is calculated using the chain rule and reflects the degree of influence of the residual connection weights on the LSTM feature extraction performance.

[0037] 3.4 SVR Layer Construction 3.4.1 SVR Model Construction and Input Logic; The core purpose of introducing SVR is to utilize its powerful regression fitting capability to transform discrete rockburst risk levels into continuous value predictions, thereby providing a quantitative correction basis for the feature weight allocation of LSTM. The input of SVR comes directly from the output of the residual connection. The residual connection has performed gradient correction and noise suppression on the temporal features of LSTM. Its output high-dimensional features can provide a more reliable regression basis for SVR and avoid redundant information in the original features from interfering with the fitting effect of SVR.

[0038] The SVR model is built around an ε-insensitive loss function. It uses a kernel function to transform high-dimensional features into a more easily fitted feature space. Its optimization objective is to minimize model complexity while keeping the deviation between predicted and true values ​​within the range of ε. The specific optimization objective formula (14) is: The constraints are formulas (15) to (17), which are as follows: ; ; , ;in Here is the weight vector of SVR, with dimension . , For the kernel function mapping dimension, this method selects the radial basis function (RBF). The value was set to 256 to balance fitting ability and computational efficiency; For SVR bias terms; This is the penalty coefficient, used to balance model complexity and fitting error. The initial value is set to 10, and it will be optimized in S4. The insensitive loss coefficient is initially set to 0.1 to control the acceptable range of prediction deviation. and These are slack variables, used to allow a small number of samples to exceed the ε bias range, thus avoiding model overfitting. For continuous labels of rockburst risk levels, the discrete labels defined in S5 are linearly mapped to... This enables SVR to perform continuous value regression; This is the kernel function mapping function; here, the RBF kernel function is used, and its expression is: , For sample index, This is the width parameter of the RBF kernel function, initially set to 0.01, used to adjust the locality of the kernel function.

[0039] To solve the constrained optimization problem described above, a Lagrange dual transformation is used to transform it into an unconstrained optimization problem, ultimately yielding the prediction formula for SVR, which is the continuously corrected value of the rockburst risk level. Specifically, as shown in formula (18): ;in The inner product form of the RBF kernel function, after simplified calculation, is expressed as follows: ; and For Lagrange multipliers, satisfy , and Its value directly determines the contribution of each sample to the prediction result; The continuously corrected values ​​output by the SVR, ranging from 0 to 1, will be passed to the attention layer of the LSTM as the core feedback signal.

[0040] 3.4.2 Bidirectional interaction between SVR, residual connection, and LSTM: The core of the deep integration of the three algorithms lies in the bidirectional interaction. SVR not only receives the feature input of the residual connection, but its output and loss also correct the parameters of the residual connection and LSTM in reverse. At the same time, the fusion features of LSTM will also feed back to optimize the parameters of SVR, forming a triple interactive closed loop.

[0041] (1) Feedback optimization of SVR → residual connection; regression loss of SVR It works in conjunction with the loss function of LSTM to update the weight matrix of the residual connections in reverse. This allows the feature correction of residual connections to simultaneously adapt to the temporal extraction of LSTM and the regression fitting requirements of SVR. The weight update formula adds a contribution term of SVR loss to the original LSTM loss feedback, specifically as formula (19): ;in The ε-insensitive loss of SVR is calculated using the following formula: , used to quantify the regression error of SVR; The loss fusion coefficient is set to 0.3 to balance the contribution weights of the LSTM temporal feature extraction loss and the SVR regression loss. The partial derivative of the SVR loss with respect to the residual connection weights is calculated using the chain rule: ,in This is derived from formula (9); This represents the sensitivity of SVR predictions to residual features. This feedback makes the residual connection a bridge between LSTM and SVR, with feature correction effects serving both algorithms simultaneously.

[0042] (2) Attention layer fusion feedback of SVR→LSTM; continuous correction value of SVR output Quantitative information containing rockburst risk levels is incorporated into the attention mechanism layer of an LSTM, allowing for dynamic adjustment of the weight allocation of the LSTM hidden states. When consecutive labels approach a certain risk level, it indicates that the features in the current LSTM feature set that are related to that level are more valuable, and the attention layer needs to strengthen the weight of these features.

[0043] The calculation process of the attention layer consists of three steps: first, fusion The attention score is calculated, then normalized to obtain the attention weights, and finally the optimized LSTM hidden state is obtained by weighting the weights. The specific formulas are as follows: Formulas (20) to (22) are as follows: ; ; ;in For the first Attention score at any given moment; Here is the attention layer weight matrix, with dimension 1. This is used to map the LSTM hidden state to a one-dimensional score; For attention layer bias terms; The weighting factor for the SVR correction value is set to 0.5 to ensure... The impact on attention scores is moderate, neither overshadowing the value of the LSTM's own features nor failing to provide effective correction. These are the normalized attention weights, summing to 1, reflecting the first... The importance of the hidden state of the LSTM at each time step; The attention-weighted LSTM hidden state represents the final temporal feature resulting from the fusion of LSTM and SVR, with a dimension of [missing information]. This feature retains the temporal dependencies captured by LSTM and incorporates the regression optimization information of SVR, significantly improving the feature representation capability.

[0044] (3) Kernel function parameter feedback of LSTM→SVR; attention-weighted LSTM hidden state It will feed back to SVR to update the width parameter of the RBF kernel function. This makes the kernel function mapping of SVR more closely match the temporal feature distribution extracted by LSTM. The update formula (23) is: ;in SVR loss pair The partial derivatives are derived from formula (18): This reflects the influence of kernel function parameters on regression error; LSTM fusion feature pairs The correction item is set to ,when and When the difference is large, it indicates that LSTM optimizes the feature distribution through the attention layer, and corresponding adjustments are needed. This feedback mechanism adapts to the new feature distribution, improving the regression accuracy of SVR. It synchronizes the SVR fitting process with the LSTM feature optimization process.

[0045] 3.5 Fully Connected Layer and Output Layer Attention-weighted fusion features Due to the high dimensionality, direct input to the output layer can easily lead to excessive model complexity and increased risk of overfitting. Therefore, a fully connected layer is introduced to perform feature dimensionality reduction and nonlinear mapping, transforming high-dimensional fused features into low-dimensional, compact feature vectors, while further exploring the complex relationships between features. The mapping formula (24) for the fully connected layer is: ;in The low-dimensional feature vector output by the fully connected layer is set to 32×1, which minimizes computational cost while preserving key information. Here is the weight matrix of the fully connected layer, with dimension 1. ; Here is the bias vector of the fully connected layer, with dimension . It is used to adjust the feature distribution and improve the fitting effect of the output layer.

[0046] The core function of the output layer is to map the low-dimensional features of the fully connected layer to the probability distribution of rockburst risk levels. A Softmax activation function is used to achieve multi-class prediction, ensuring the output meets probability normalization requirements and facilitating direct interpretation of the likelihood of each risk level. The output formula (25) is: ;in The output layer weight matrix has dimensions of . 4 corresponds to the four rockburst risk levels defined in S5, from Level I to Level IV; The output layer bias vector has a dimension of . ; Let be the predicted probability distribution vector, in the form of ;in For the current sample to belong to the first The probability of a level 1 rockburst risk. Corresponding to Level I Corresponding to Level II Corresponding to Level III Corresponding to Level IV, and ; The normalized exponential function is expressed as follows: ; The linear mapping result of the output layer. The The function transforms the linear mapping result into a probability distribution, making the prediction results more interpretable.

[0047] 3.6 Model Training Process The core of model training is to achieve collaborative parameter optimization of three algorithms: improved LSTM, residual connections, and SVR. Forward propagation generates prediction results and calculates the total loss. Then, backpropagation distributes the loss gradient to the parameters of each algorithm, iteratively updating until the model converges. The specific steps are as follows: 1. Parameter Initialization Phase: Initialize all trainable parameters in the model to ensure a reasonable starting point for training. Improved LSTM parameters include embedding layer weights. Bias and the weight matrix of each gate control , , , and bias vector , , , Specific parameters of the second-layer LSTM , All parameters are initialized using the Xavier method to adapt the parameter distribution to the input range of the activation function; the weights of the residual connections... and bias The data was initialized using a normal distribution with a mean of 0 and a variance of 0.01; the parameters of the SVR were... , Initialized to 0, penalty coefficient Kernel function parameters Insensitive loss coefficient Assign values ​​according to the initial values ​​set above; weights of the attention layer. and bias Uniform initialization is used, with a range of [-0.05, 0.05]. The weights and biases of the fully connected layers and the output layer are also initialized using Xavier to ensure the stability of the feature distribution of each layer.

[0048] 2. Forward Propagation Phase: Forward propagation is the process of features flowing and fusing between algorithm modules. It is executed strictly according to the architectural order, ensuring that the output of each step serves as the input for subsequent modules. First, the training set data divided by S2 is... The input model is embedded and mapped using formula (2) to obtain high-dimensional features. Subsequently, Input the first LSTM layer and generate temporal features through gating operations in formulas (3) to (8). ; and The common input residuals are connected, and the corrected features are obtained by formula (9). ; As input to SVR, continuous correction values ​​are output through regression operations of formulas (14) to (18). ;at the same time, Input the second LSTM layer, and extract the deep features using formulas (10) to (12) to obtain ; and The collaborative input attention layer generates fused features through weight adjustments according to formulas (20) to (22). ; The input fully connected layer is dimensionality reduced using formula (24) to obtain ;at last, The predicted probability distribution is output through the Softmax mapping of formula (25). Throughout the process, the outputs of the three types of algorithms are interdependent, forming a feature flow closed loop of LSTM → residual connection → SVR → LSTM.

[0049] 3. Loss Calculation Stage: Loss calculation is the core of evaluating the model's predictive performance. It requires calculating the specific losses for each of the three algorithms separately, then merging them into a total loss to ensure that the optimization objectives of each algorithm are consistent. First, the cross-entropy loss of the improved LSTM is calculated. This is used to quantify the difference between the predicted probability and the true label, and the formula is: ,in The total number of samples in the training set. For the first The true labels for each sample are encoded using one-hot encoding, such as the corresponding labels for a Level I rockburst. ; For this sample, the corresponding number First, predict the probability of level-one risk; then calculate the ε-insensitive loss of SVR. Used to quantify continuous correction values With continuous tags The deviation is given by the formula: Finally, the two types of losses are combined according to their weights to form the total loss. ,in The loss fusion coefficient ensures that the temporal feature extraction of LSTM is as important as the regression optimization of SVR, avoiding a single loss dominating the model training direction.

[0050] 4. Backpropagation Stage: Backpropagation is the core of parameter optimization. It calculates the partial derivatives of the total loss with respect to each parameter using the chain rule, and updates the parameters in the following order: output layer → fully connected layer → attention layer → SVR / second LSTM layer → residual connection → first LSTM layer → embedding layer, achieving coordinated optimization of the three types of algorithms. First, it calculates the total loss with respect to the output layer weights. Bias and fully connected layer weights Bias The partial derivatives are updated using gradient descent; then, the parameters of the attention layer are calculated. , The partial derivatives are updated and fed back to the SVR and the second LSTM; then, for the SVR, the total loss with respect to the Lagrange multipliers is calculated. , and bias The partial derivatives are used to update the kernel function parameters via formula (23). For the second LSTM layer, calculate and update the partial derivatives of its gating parameters; then, update the weights of the residual connections using formula (19). and bias This update integrates the gradient contributions of the LSTM loss and SVR loss; finally, the parameters of the first LSTM layer and the embedding layer are updated, completing one backpropagation. Throughout the process, the parameter updates of each algorithm are interconnected. The loss gradient of SVR affects the parameters of the residual connection and LSTM, and the feature feedback of LSTM affects the parameters of SVR, ensuring that the three types of algorithms form an optimization closed loop.

[0051] 5. Iterative convergence phase: Repeat the forward and backward propagation processes, updating the parameters with a batch of training set samples in each iteration, until the total loss is reached. The convergence condition is met. The convergence condition is set as the total loss decreasing by less than [a certain value] over 10 consecutive epochs. The iterations may continue until the maximum number of iterations set by S4 is reached (initially within the range of 100-500). During the iteration process, after each epoch, the model performance is evaluated using the validation set partitioned by S2. If the validation set performance does not improve for 10 consecutive epochs, an early stopping strategy is triggered to prevent overfitting of the model, while the model corresponding to the current optimal parameters is saved.

[0052] 3.7 Model Application Process 1. New Project Data Acquisition and Preprocessing: Following the rockburst influencing factor categories and acquisition methods specified in S1, acquire the raw data for the project to be predicted, including geostress parameters obtained from field stress measurements, rock mechanics parameters obtained from laboratory tests, engineering geological condition parameters extracted from the engineering survey report, and construction disturbance parameters obtained from construction records. Subsequently, strictly follow the preprocessing procedure in S2 to process the raw data: use mean imputation or median imputation to handle missing values, use the 3σ principle or box plot method to detect and correct outliers, and then use the Z-score standardization method to convert the data into standardized values. To ensure that the preprocessing method is completely consistent with the training set, and to avoid the impact of data distribution differences on the prediction results.

[0053] 2. Collaborative Feature Extraction and Correction: This involves extracting and correcting the standardized data to be predicted. Input the trained model and start the collaborative feature processing flow of the three algorithms. First, obtain the high-dimensional features through the embedding layer formula (2). Subsequently, the first LSTM layer extracts temporal features using formulas (3) to (8). The residual connection is obtained by correcting it using formula (9). This both alleviates gradient vanishing and provides high-quality features for SVR; SVR receiver Then, the continuous correction value is output using formula (18). This value quantifies the tendency of the rockburst risk level of the sample to be predicted; at the same time, Input the second LSTM layer, and further mine deep temporal features through formulas (10) to (12). Attention layer fusion and By dynamically adjusting the feature weights using formulas (20) to (22), key features related to the current risk level are strengthened, resulting in the final fused features. Throughout the process, the three types of algorithms work together according to the trained parameters, and feature extraction, correction, and weight optimization are completed in one go, ensuring the accuracy of the fused features.

[0054] 3. Risk Level Prediction and Output: Final Fusion Features The dimension reduction obtained by formula (24) of the fully connected layer is Then, the probability distribution of rockburst risk level of the sample to be predicted is generated by using the Softmax function of the output layer formula (25). The risk level corresponding to the maximum value in the probability distribution is selected as the final prediction result, i.e., if... The maximum predicted level is I, with no rockburst. The maximum predicted level is II, a weak rockburst; The maximum predicted level is Level III, a moderate rockburst. The maximum predicted risk level is IV, a strong rockburst. Simultaneously, it outputs the probability values ​​for each risk level, providing richer reference information for engineering decisions. For example, if... , If the sample is between a weak and a moderate rockburst, then more cautious protective measures are required.

[0055] 3.8 Explanation of Core Role and Integration Value 3.8.1 The core function of each algorithm; 1. The core role of the improved LSTM: As the core temporal feature extractor of the model, it is specifically designed for the dynamic formation mechanism of rockburst risk. The occurrence of rockburst is closely related to temporal processes such as the accumulation of ground stress and the continuous effects of construction disturbances. Traditional models struggle to capture these time-varying dependencies. The improved LSTM, through a two-layer stacked structure and gating mechanism, can accurately remember early key features such as the initial ground stress state and forget redundant information such as occasional construction fluctuations. It extracts the synergistic patterns of different influencing factors over time, providing core temporal feature support for subsequent feature correction and risk prediction. Its two-layer stacked design further enhances the ability to extract deep features, enabling the model to uncover more complex temporal correlations.

[0056] 2. The core role of residual connections: to solve the gradient vanishing problem in deep neural networks, and simultaneously build an improved feature bridge between LSTM and SVR. In rockburst prediction, gradient vanishing in deep LSTMs prevents early temporal features, such as stress states in the early stages of construction, from being propagated to deeper layers, affecting the model's ability to capture long-term dependencies. Residual connections, through direct paths, allow embedded layer features to be superimposed onto the LSTM hidden state, enabling gradients to propagate directly back to shallower layers, ensuring the complete transmission of temporal features; simultaneously, their output corrects features... It retains the temporal information of LSTM while suppressing noise, providing high-quality regression input for SVR and avoiding redundant information in the original features from interfering with the fitting effect of SVR. In addition, the parameter update of the residual connection integrates the loss feedback of LSTM and SVR, making its feature correction effect adaptable to the needs of both algorithms, and becoming a key link for synergistic integration.

[0057] 3. The core function of SVR: To compensate for the insufficient sensitivity of LSTM to discrete risk levels by using regression capabilities, providing a quantitative correction signal. Rockburst risk levels are discretely classified (Levels I-IV). When LSTM directly outputs classification probabilities, its ability to distinguish boundary samples, such as those between weak and moderate rockbursts, is weak. SVR transforms discrete levels into continuous value regression, outputting... The risk tendency of quantifiable samples provides a precise basis for adjusting the attention layer of the LSTM. As the model approaches a specific consecutive label, the attention layer strengthens the weights of the corresponding features, improving the classification accuracy of boundary samples. Simultaneously, the regression loss of SVR is fed back to the residual connections and LSTM, enabling the model to extract temporal features while also considering the quantitative fitting of risk levels, thus improving prediction accuracy and robustness.

[0058] 3.8.2 The Core Value of Three-Algorithm Fusion The deep integration of the three types of algorithms is not a simple superposition of modules, but rather a closed loop formed through bidirectional interaction: temporal feature extraction, feature correction, quantization feedback, and weight optimization. This solves the inherent shortcomings of a single algorithm in rockburst prediction. It overcomes the problems of gradient vanishing and unreasonable key feature weight allocation in traditional LSTM: residual connections ensure the transmission of deep features, and the quantization correction of SVR enables the attention layer to accurately identify key features. It makes up for the inability of SVR to capture time-series dependencies: the time-series features provided by LSTM make the regression fit of SVR more closely match the dynamic formation process of rockburst, avoiding the prediction bias caused by static regression. The model's generalization ability and robustness have been improved: the synergistic optimization of the three types of algorithms enables the model to adapt to different geological conditions and construction scenarios in various projects, and significantly enhances its ability to handle noisy data and boundary samples.

[0059] In rockburst prediction scenarios, this integrated value is specifically manifested in the fact that it can accurately capture the temporal variation patterns of factors such as ground stress and construction disturbance, and improve the accuracy of risk level classification through quantitative correction. It has a significant advantage, especially in distinguishing boundary rockburst risks under complex geological conditions, such as weak-to-medium rockbursts and medium-to-strong rockbursts, providing more reliable predictive support for engineering construction safety.

[0060] S4: Model Training and Hyperparameter Optimization Based on the training and validation sets output by S2, and combined with the collaborative fusion model constructed by S3, parameter collaborative optimization training is carried out to ensure that the model achieves optimal performance in temporal feature extraction, regression correction, and weight allocation. The specific implementation details are as follows: 1. Hyperparameter initialization; clearly define all hyperparameters to be optimized and their value ranges, as shown in the example below: Model structure parameters: LSTM unit count: 64, 128, 256; fully connected layer dimension: 16, 32, 64; training parameters: learning rate: 0.001, 0.005, 0.01; number of iterations: 100, 200, 300, 400, 500; batch size: 8, 16, 32; LSTM regularization coefficient: 0.0001, 0.0005, 0.001; SVR parameters: penalty coefficient C: 1, 5, 10, 20; kernel function parameter γ: 0.001, 0.01, 0.1, 1; insensitive loss coefficient ε: 0.05, 0.1, 0.2. 2. Collaborative training implementation: The loss function adopts the total loss defined in S3. , λ=0.3, where For LSTM cross-entropy loss, The SVR is used for ε-insensitive loss; the Adam optimizer is selected, with momentum parameters β1=0.9, β2=0.999, and weight decay coefficient set to 0.0001 to avoid parameter oscillation during training; the training process is as follows: forward propagation to generate prediction results → calculate total loss → backpropagation to distribute gradients → collaborative parameter update. The parameter update order is: output layer → fully connected layer → attention layer → SVR / second LSTM layer → residual connection → first LSTM layer → embedding layer, to ensure that the parameters of each algorithm are optimized synchronously. 3. Hyperparameter optimization: All hyperparameter combinations are traversed using a grid search method. Each combination is trained 3 times and the average performance is taken. The average of the accuracy on the validation set plus the F1 score is used as the optimization objective, with each weighted at 0.5, to select the optimal hyperparameter combination. 4. Early Stopping and Model Saving: Enable the early stopping strategy and monitor and validate the ensemble loss. If the total loss decreases by less than [a certain percentage] for 10 consecutive epochs, [further action will be taken]. If the performance on the validation set does not improve, stop training immediately; save the model with the best overall performance on the validation set as the base model for subsequent evaluation and prediction.

[0061] This step outputs the best co-fusion model after training, which includes all the optimal parameters for LSTM, residual connections, and SVR. This model will be used for performance evaluation in S6 and risk prediction in S7.

[0062] S5: Rockburst Risk Level Label Code To simultaneously adapt to both LSTM classification and SVR regression tasks, rockburst risk levels need to be dual-labeled to ensure that the labels are consistent with the model output requirements. Specific implementation details are as follows: 1. Risk Level Classification: Rockburst risk is strictly classified into four levels according to the rockburst classification standard, with clear criteria for each level: Level I (No Rockburst): No signs of rockburst, rock is stable after excavation, no spalling or ejection; Level II (Weak Rockburst): Local rock spalling, occasional small particle ejection, no obvious vibration; Level III (Moderate Rockburst): Moderate rock spalling, frequent particle ejection, accompanied by slight vibration; Level IV (Strong Rockburst): Large-scale rock spalling or throwing, strong vibration, affecting construction safety.

[0063] 2. Label Encoding Rules: Discrete labels, adapted for LSTM classification: a combination of integer encoding and one-hot encoding is used, Level I = 0 (one-hot: [1,0,0,0]), Level II = 1 ([0,1,0,0]), Level III = 2 ([0,0,1,0]), Level IV = 3 ([0,0,0,1]); Continuous labels, adapted for SVR regression: discrete labels are linearly mapped to continuous values ​​in the interval [0,1], Level I = 0, Level II = 0.33, Level III = 0.66, Level IV = 1.0, ensuring that SVR can fit continuous values; 3. Label and dataset association: The label for each time series sample is determined by the actual rockburst level corresponding to the last data of the time series sample, that is, the risk level corresponding to 10 consecutive sets of influencing factor data, based on the actual rockburst situation at the end of the time period, to ensure the temporal matching of features and labels.

[0064] This step outputs a complete dataset with dual labels: a training set with discrete and continuous labels, a validation set with both discrete and continuous labels, and a test set with both discrete and continuous labels. The discrete labels are used for LSTM loss calculation, and the continuous labels are used for SVR loss calculation.

[0065] S6: Model Performance Evaluation and Optimization The model is comprehensively evaluated from both classification and regression performance using the test set output by S2 and the labels encoded by S5. Targeted optimizations are then performed to address any shortcomings. Specific implementation details are as follows: 1. Evaluation Metric Calculation: Classification Performance Metrics: Accuracy: Overall classification correctness; Precision: The proportion of correctly predicted values ​​at each level; Recall: The proportion of actual samples correctly predicted at each level; F1 Score: The harmonic mean of precision and recall; Each metric is calculated according to its corresponding formula, and the results are rounded to two decimal places; Regression Performance Metrics: Mean Squared Error: MSE, quantifies the deviation between the continuously predicted SVR values ​​and the true continuous labels; Coefficient of Determination: To measure the goodness of fit of SVR, ≥0.8 indicates a good fit; 2. Performance Standards: Set comprehensive performance thresholds that must simultaneously meet the following requirements: accuracy ≥ 85%, F1 score ≥ 80%. ≥0.8, total loss ≤0.2; 3. Targeted optimization measures: If classification performance is substandard, such as high misclassification rates for Level III and IV: adjust the β coefficient of the LSTM attention layer, the weight of the SVR correction value, or increase the number of LSTM units and extend the time series length T; if regression performance is substandard, such as MSE≥0.05: optimize the C, γ, and ε parameters of the SVR, or change the kernel function type, such as a polynomial kernel; if overall performance is substandard: return to S1 to supplement time series samples corresponding to the risk level, such as adding Level IV rockburst samples, and repeat the S2 to S6 process for retraining; 4. Final Model Determination: The model that passes all the thresholds is the final collaborative fusion prediction model. If it still fails to meet the standards after multiple optimizations, the model architecture of S3 needs to be re-examined, such as adjusting the weight coefficients of the residual connections.

[0066] This step outputs the final version of the rockburst risk classification prediction model. This model has excellent time series feature extraction, regression correction and classification prediction capabilities, and can be directly used for actual engineering prediction.

[0067] S7: Application of Rockburst Risk Classification and Prediction Based on the final model, rockburst risk prediction for new projects is carried out according to standardized procedures. The application and iteration process of the final model prediction is as follows: Figure 4 As shown. Specific implementation details are as follows: 1. New project data acquisition: According to the influencing factor categories and acquisition standards in S1, acquire real-time data of the project to be predicted, with the acquisition frequency consistent with the training data; for missing individual data, process them in real time according to the cleaning rules in S2 to avoid data delays affecting the prediction; 2. New Data Preprocessing: Process the new data according to the S2 process: Construct time series samples, with 10 consecutive data sets as one sample; Z-score standardization, using μ and σ from the training set to avoid data distribution shift; The processed data format is (M, 10, 12), where M is the number of samples to be predicted; 3. Collaborative prediction execution: Input the preprocessed data to be predicted into the final model and start the collaborative prediction process: LSTM extracts temporal features → residual connection correction → SVR outputs continuous correction values ​​→ attention layer optimizes weights → fully connected layer reduces dimensionality → outputs probability distribution; 4. Interpretation and Output of Prediction Results: Core Result: The risk level corresponding to the maximum value in the probability distribution is selected as the final level. For example, if the maximum probability is 0.75, corresponding to Level II, then the prediction is a weak rockburst. Auxiliary Information: Output the probability value and SVR continuous correction value for each level. If the maximum probability is <0.7 (boundary sample), it is necessary to mark it as "boundary risk" and give the tendency level, such as "prone to Level II, monitoring needs to be strengthened". Output Format: Generate a prediction report, including continuous correction values, final risk level and protection recommendations, such as Level I does not require special protection, and Level IV requires suspension of construction and pressure relief measures.

[0068] This step outputs a rockburst risk classification prediction report for the new project, providing a direct basis for decision-making on construction safety.

[0069] S8: Prediction Validation and Model Iteration Optimization To ensure the model maintains high prediction accuracy in long-term applications, a closed-loop optimization mechanism of prediction-verification-iteration needs to be established. The model should be corrected based on the actual rockburst conditions of the new project. The specific implementation details are as follows: 1. On-site Verification and Implementation: Tracking Period: From the start of construction of the project to be predicted to 72 hours after the completion of the predicted section, ensuring coverage of the complete period during which rockbursts may occur; Verification Data Recording: Based on the S5 risk level judgment criteria, record the actual rockburst situation in real time, including whether a rockburst occurred, the rockburst level (e.g., large-scale rock spalling observed on-site is classified as Level IV), the occurrence time, and corresponding influencing factor data such as the maximum principal stress and blasting intensity at the time of the rockburst, forming an actual rockburst dataset with a format that corresponds one-to-one with the predicted samples (M, 10, 12+1), where the last column is the actual risk level label; Verification Accuracy Quantification: Calculate the degree of agreement between the predicted results and the actual results. The core indicator is the actual rockburst level accuracy rate: the percentage of samples that correctly predicted whether a rockburst occurred and its level; Major Risk Miss Rate: the proportion of Level IV rockbursts predicted as Level III or below must be ≤5%; Over-warning Rate: the proportion of no rockbursts predicted as Level II or above must be ≤10%. 2. Result Comparison and Misjudgment Analysis: Sample-by-sample comparison: The prediction results of S7 were matched one by one with the actual rockburst data, and misjudged samples were marked, such as predicted as level II but actually being level III, or predicted as no rockburst but actually being a weak rockburst; Misjudgment type classification: Boundary misjudgment: Misjudgment caused by continuous correction values ​​between two levels, mostly due to insignificant changes in time series characteristics; Feature missing misjudgment: Special factors not covered in the training data exist in the new project, such as sudden groundwater inrush, which the model cannot identify; Parameter adaptation misjudgment: The model parameters are not adapted to the geological conditions of the new project, such as the parameter difference between high-stress soft rock areas and hard rock areas; Root cause localization: Combining the evaluation indicators of S6, if the boundary misjudgment ratio is ≥60%, it points to the attention layer β coefficient or SVR parameter; if the feature missing misjudgment ratio is high, it is necessary to supplement the influencing factors; if the parameter adaptation misjudgment is prominent, it is necessary to optimize the model's generalization ability.

[0070] 3. Model Iterative Optimization Measures: Sample Expansion Optimization: After preprocessing actual rockburst data according to the S2 standard, add it to the training set and re-execute the S4 co-training to improve the model's adaptability to new geological conditions; Parameter Fine-tuning Optimization: Targeted parameter adjustments based on the causes of misjudgments: Boundary Misjudgment: Adjust the attention layer β coefficient, such as from 0.5 to 0.6, and the SVR ε parameter, such as from 0.1 to 0.08, to enhance the discriminative power of boundary features; Over-warning / Under-judgment: Adjust the LSTM regularization coefficient, such as from 0.0005 to 0.001, and the SVR penalty coefficient C... For example, adjusting from 10 to 15 balances the sensitivity and robustness of the model; architecture adaptation optimization: if the geological conditions of the new project are special, such as deep high stress or strong groundwater, return to S3 to adjust the model architecture: such as increasing the number of residual connection layers, changing the SVR kernel function to a hybrid kernel (RBF + polynomial kernel), or adding a feature screening module after the embedding layer; emergency optimization mechanism: if the false negative rate of major risks is >5%, immediately suspend the prediction application of the project, prioritize supplementing extreme samples of the corresponding risk level, such as strong rockburst time series data, retrain, and pass the rigorous evaluation of S6 before it can be resumed.

[0071] 4. Model Updates and Maintenance: Regular Updates: Set a model update cycle, such as every 3 engineering predictions completed or every 6 months, integrate the actual rockburst data of all engineering projects, re-execute the S2 to S6 processes, generate a new version model, and replace the old model; Version Management: Record the model parameters, training data, evaluation indicators, and applicable engineering types such as hard rock tunnels and soft rock mines for each version, and establish a model archive to facilitate the selection of an appropriate version for subsequent projects based on geological conditions; Anomaly Monitoring: Monitor the predicted probability distribution of the model in real time. If the maximum probability of 5 consecutive samples is <0.6, trigger an alert and prompt technicians to check the data acquisition quality or initiate parameter fine-tuning.

[0072] Through continuous field validation and targeted optimization, the model's performance is dynamically improved, ensuring its predictive reliability under different geological conditions and construction scenarios, thus forming a closed-loop management system covering the entire lifecycle of data acquisition, model training, prediction application, and validation optimization.

[0073] In some embodiments, a quantum convolutional neural network (QCNN) algorithm is incorporated to compensate for the shortcomings of traditional algorithms in capturing fine local features. Leveraging the superposition and entanglement properties of quantum states, QCNN can efficiently uncover the coupling correlations between factors such as ground stress and construction disturbances within a local spatiotemporal range, providing more accurate local feature support for subsequent feature correction and temporal extraction. Its integration with existing algorithms follows a closed loop of local feature enhancement and global collaborative optimization, achieving a bidirectional flow of quantum features → classical features → quantum parameter feedback.

[0074] I. Construction of Quantum Convolutional Neural Network (QCNN) Model 1.1 Model Hierarchy Construction (1) Quantum feature mapping layer: classical features → quantum state; high-dimensional classical features output by the embedding layer. Dimension m=128 For the first The standardized values ​​of each feature are mapped to quantum states. The feature representation space is expanded using quantum superposition. The formula is: ; : Number of qubits, because Therefore Used for building Wichlbert space; Initial quantum state, default ,Right now The ground state of 1 qubit; : Parameterized quantum mapping gate, composed of rotation gates constitute, This maps classical eigenvalues ​​to quantum gate rotation angles. The mapped quantum state, dimension , in the form of , For quantum amplitude, satisfying , To calculate the ground state; Quantum rotation gate matrix.

[0075] Calculation process: For Each feature Calculate the rotation angle ,scope ;right Each qubit is applied Gate; Output superposition state This completes the mapping from classical features to quantum states.

[0076] (2) Quantum convolutional layer: Extracts local correlation features; captures the local coupling relationship between adjacent features in the quantum state through quantum convolution kernel, and strengthens the local synergistic effect of rockburst influencing factors, such as the coupling between in-situ stress and rock mechanical parameters in a certain area; Formula: , ; ; A quantum convolution gate consists of a 2-qubit controlled gate (CNOT) and a rotation gate. For quantum convolution kernel parameters, Number of convolution kernels; : No. The quantum state output by a quantum convolution kernel; The superposition of all quantum convolution kernel outputs is achieved through quantum multiplexing, with dimensionality... .

[0077] Calculation process: Initialization Individual quantum convolution kernels to Each core corresponds to a set of CNOT gates and Gate combination parameters; each convolution kernel pair Local feature extraction is performed: entanglement of adjacent qubits is established using CNOT gates to simulate local feature correlations. Adjusting feature weights using gates; utilizing quantum multiplexing to The output states of each convolution are superimposed to obtain Preserve multi-scale local features.

[0078] (3) Quantum Pooling Layer: Dimensionality Reduction and Feature Aggregation; Dimensionality reduction of features is achieved through quantum measurement, preserving key local features and reducing subsequent computational complexity; Formula: ; , ; ; Quantum pooling gates employ a local maximum measurement strategy. These are pooling parameters. Number of pooled windows; A set of quantum measurement operators For projection measurement; The classic feature vector output after measurement, with dimensions... ; : No. The measurement results of each pooling window represent the expected probability of local features.

[0079] Calculation process: For Applying quantum pooling gates ,Will The qubits are grouped into adjacent pairs, totaling... One group is formed, and the remaining 1 bit is directly reserved; a projection measurement is applied to each group of qubits. ,get Output classic feature vectors Dimensions from Down to This achieves efficient dimensionality reduction.

[0080] (4) Quantum Fully Connected Layer: Classical Feature Fusion and Output; This layer fuses the classical features obtained from quantum pooling with quantum state information to output the final QCNN local feature vector, which is used for interaction with other algorithms; Formula: ; ; QCNN fully connected layer weight matrix QCNN outputs feature dimensions; : QCNN fully connected layer bias vector; The QCNN activation function uses LeakyReLU to avoid gradient vanishing. The classic local feature vector output by QCNN, with dimensions... ; : Linear mapping result of fully connected layer.

[0081] Calculation process: Weights of fully connected layers Initialize as a normal distribution bias Initialize to 0; compute linear mappings It incorporates quantum pooling features; and outputs the LeakyReLU activation function. It preserves the nonlinear correlation of local features and avoids the complete loss of negative features.

[0082] II. Two-way interaction design between QCNN and other algorithms 2.1 Core principle of interaction: All interactions are bidirectional closed loops. The output features of QCNN serve as inputs to other algorithms, and the losses / features of other algorithms are fed back to QCNN to optimize its parameters.

[0083] 2.2QCNN↔Residual Connections: Bidirectional Interaction Between Local Features and Gradient Correction (1) Forward interaction: QCNN local feature enhancement residual correction; the corrected features of residual connections are integrated into QCNN local features, improving the "local-global" fusion of features and providing more accurate low-noise features for LSTM deep temporal extraction; Formula: ;in Weight matrix for the interaction between residual connections and QCNN; Residual correction features incorporating QCNN features replace the original solution. .

[0084] Interaction process: Local features output by QCNN Through weight Mapped to the residual connection feature space; and associated with LSTM temporal features. Embedded layer features Synergistic integration, generation ; Inputting a second-layer LSTM enables temporal feature extraction to simultaneously incorporate global temporal dependencies and local coupling relationships.

[0085] (2) Reverse interaction: The residual connection loss is fed back to optimize the QCNN parameters; the weight update loss of the residual connection is backpropagated to the QCNN to adjust its convolutional kernel and fully connected layer parameters, so that the local feature extraction adapts to the residual correction requirements; Formula: ; ; Total loss function QCNN loss , These are local feature labels, annotated from local rockburst observation data; , Loss fusion coefficient, balancing temporal, regression, and local feature losses; , : The partial derivatives of the total loss with respect to the QCNN fully connected weights and convolutional kernels.

[0086] Interaction process: LSTM cross-entropy loss Backpropagation to residual connection weights ,pass Associated with QCNN features Combined with QCNN's own loss Calculate the total loss with respect to the QCNN parameters , The gradient is used to update the QCNN parameters, making local feature extraction more aligned with the temporal-local fusion requirements of residual connections.

[0087] 2.3QCNN↔SVR: Bidirectional Interaction Between Local Features and Regression Correction (1) Positive interaction: QCNN local features improve SVR regression accuracy; the input of SVR incorporates QCNN local features, strengthening the quantitative fitting of local factors that play a dominant role in rockburst risk level, such as the influence of the local coupling between joint density and blasting intensity in a certain area on the risk level; formula: ; ;in This represents the weight matrix for the interaction between SVR and QCNN. The final input features for SVR are fused with residual features and QCNN local features.

[0088] Interaction process: QCNN local features pass Mapped to the SVR feature space, and with residual correction features Fusion; Fusion characteristics Input SVR, through RBF kernel function Calculate sample similarity and output continuous correction values. The incorporation of local features enables SVR to accurately quantify the contribution of "local factor coupling" to the risk level and improve the fitting accuracy of continuous values ​​for boundary samples.

[0089] (2) Backward interaction: The SVR regression loss is fed back to optimize the QCNN pooling parameters; the SVR regression loss is backpropagated to the QCNN to adjust its pooling parameters. This makes local feature extraction more suitable for the regression fitting requirements of SVR; Formula: ; ;in The partial derivative of the SVR loss with respect to the QCNN pooling parameters; The partial derivative of the QCNN loss with respect to its own pooling parameters.

[0090] Interaction process: SVR regression loss pass Associated with fusion input This is then passed on to the QCNN features. Combined with QCNN's own loss Calculate the total loss with respect to pooling parameters gradient; update This allows quantum pooling to more accurately preserve local features that play a key role in regression fitting.

[0091] 2.4QCNN↔ Improved LSTM: Bidirectional interaction between local and temporal features (1) Positive interaction: QCNN local features enhance LSTM temporal extraction; QCNN local features are integrated into the LSTM attention layer, dynamically adjusting the temporal feature weights, making LSTM pay more attention to temporal changes related to local coupling; Formula: ; ; ; ;in This represents the weight matrix for the interaction between LSTM and QCNN; The final input features of the attention layer are fused with LSTM temporal features and QCNN local features; Attention score.

[0092] Interaction process: QCNN local features pass Mapped to the LSTM feature space, and combined with deep temporal features Fusion; Fusion characteristics With SVR correction value The attention layer is input together, and the attention weights are calculated. Weighted It also includes temporal dependency and local coupling information, improving LSTM's ability to capture complex temporal evolution.

[0093] (2) Backward interaction: The LSTM temporal loss is fed back to optimize the QCNN mapping gate parameters; the LSTM cross-entropy loss is backpropagated to the QCNN to adjust its quantum mapping gate parameters. This makes quantum feature mapping more suitable for the needs of temporal feature extraction; Formula: ; ;in The partial derivative of the LSTM loss with respect to the QCNN mapping gate parameters; The partial derivative of the QCNN loss with respect to its own mapping gate parameters.

[0094] Interaction process: LSTM cross-entropy loss Through attention weight Associated with fusion input This is then passed to the QCNN quantum state. Combined with QCNN's own loss Calculate the total loss with respect to the mapping gate parameters. gradient; update This enables quantum feature mapping to capture classical features related to temporal evolution more accurately, thereby improving the local-temporal collaborative extraction effect.

[0095] III. QCNN Model Training Process 3.1 Parameter Initialization; QCNN-Specific Parameter Initialization: Quantum Convolution Kernel ( Each dimension ); Quantum pooling parameters , One; weights of the fully connected layer ; bias Mapping gate rotation angle The initial values ​​are calculated from the embedding layer features. Interaction weight initialization: ; ; .

[0096] 3.2 Forward Propagation; Embedded Layer Output Inputting into a QCNN, the algorithm passes through quantum mapping, convolution, pooling, and fully connected layers to output local features. ; Input residual connections, SVR, and LSTM attention layers respectively, and fuse them with features from other algorithms to generate... , , The subsequent process follows the original model logic: LSTM deep extraction, SVR regression, attention weighting, and fully connected output, ultimately yielding the predicted probability distribution. .

[0097] 3.3 Loss Calculation; The total loss function is updated as follows: The batch sample QCNN loss is: , For batch size, For the first The true label of local features of each sample.

[0098] 3.4 Backpropagation; Calculate the gradient of the total loss with respect to each algorithm parameter; Update parameters in the order of output layer → fully connected layer → attention layer → SVR / LSTM / QCNN → residual connection → embedding layer; QCNN parameters and interaction weights are updated synchronously to ensure collaborative optimization.

[0099] 3.5 Iterative convergence; the convergence condition remains unchanged, and the total loss decreases by less than 1 / 2 for 10 consecutive epochs. Or reach the maximum number of iterations; early stopping strategy simultaneously monitors QCNN loss. ,like If the model does not decrease for 10 consecutive epochs, an early stop is triggered and the optimal model is saved.

[0100] IV. QCNN Model Application Process New project data preprocessing: Obtain standardized data according to the S2 workflow. Generated through the embedding layer QCNN collaborative feature extraction: Input QCNN, output local features Multi-algorithm collaborative prediction: Incorporate residual connections to generate Input the second LSTM layer; Integrate SVR input to generate Input attention layer; Integrating into the LSTM attention layer to generate Input: Fully connected layer; Output: Prediction result: After passing through the fully connected layer and the Softmax function, the output is the probability distribution of rockburst risk level, which includes quantitative information on the contribution of local features of QCNN.

[0101] V. The Core Role and Integration Value of QCNN Core functions: Enhancement of local nonlinear features: By leveraging the properties of quantum entanglement and superposition, it captures the local coupling correlation of rockburst influencing factors, making up for the shortcomings of traditional algorithms that focus on the global rather than the local; Enhancement of boundary sample differentiation: In collaboration with SVR and LSTM, it accurately identifies local differences in boundary samples, reducing the misjudgment rate in scenarios such as weak-medium and medium-strong rockbursts; Expansion of feature expression space: It enriches feature expression through quantum state mapping, improving the model's generalization ability.

[0102] Integration Value: Building a Collaborative Closed Loop: Forming a triple collaboration of local, temporal, and regression, with four types of algorithms interacting bidirectionally, covering the entire process of rockburst risk formation; Adapting to Extreme Working Conditions: Efficiently handling high-dimensional nonlinear features, adapting to different geological conditions, and reducing the risk of misjudgment in extreme scenarios; Enhancing Engineering Practicality: Standardized collaborative processes are adapted to real-time monitoring, and prediction results provide dual references for global risk and local key points.

Claims

1. A rockburst risk classification and prediction method based on an improved LSTM neural network, characterized in that, Includes the following steps: S1: Collect data on rockburst influencing factors to form a raw dataset of rockburst influencing factors; S2: Clean the original dataset of rockburst influencing factors, construct time-series samples, and standardize the data to obtain a standardized dataset. Divide the standardized dataset into training set, validation set, and test set. S3: Construct a collaborative fusion neural network model; S4: Based on the training set and validation set, the collaborative fusion neural network model is trained using the total loss function, and the optimal model is determined through hyperparameter optimization and early stopping strategy; S5: The rockburst risk level is encoded with dual labels to obtain discrete labels adapted to the improved LSTM classification and continuous labels adapted to SVR regression. S6: Based on the test set, discrete labels and continuous labels, evaluate the classification performance and regression performance of the optimal model, and obtain the final collaborative fusion prediction model by judging the threshold. S7: Collect data on rockburst influencing factors for the new project, preprocess the data according to the standard in step S2, input the data into the final collaborative fusion prediction model, and output the rockburst risk classification prediction results for the new project.

2. The rockburst risk classification and prediction method based on an improved LSTM neural network according to claim 1, characterized in that, The S3 collaborative fusion neural network model consists of an embedding layer, an improved LSTM layer, a residual connection algorithm module, an SVR module, an attention layer, a fully connected layer, and an output layer. The embedding layer is used to map the features of the standardized dataset to a high-dimensional feature space and output high-dimensional features. The LSTM layer is improved to a two-layer stacked structure to extract the temporal dependency features of rockburst influencing factors; The residual connection algorithm module is built between two LSTM layers. It simultaneously receives the temporal features output from the first LSTM layer and the high-dimensional features output from the embedding layer, and outputs the corrected features. The SVR module receives the corrected features and outputs continuously corrected values ​​through regression operations. The attention layer receives the deep temporal features output from the second LSTM layer and the continuously corrected values ​​output from the SVR module, fuses and calculates the attention weights, and optimizes the deep temporal features by weighting them to obtain the fused features. The fully connected layer performs dimensionality reduction mapping on the fused features, and the output layer outputs the probability distribution of rockburst risk level through the Softmax function.

3. The rockburst risk classification and prediction method based on an improved LSTM neural network according to claim 1, characterized in that, It also includes S8: tracking the actual rockburst situation of new projects to form an actual rockburst dataset, comparing the prediction results with the actual rockburst data, and achieving iterative optimization of the model through sample expansion, parameter fine-tuning or architecture adaptation.

4. The rockburst risk classification and prediction method based on an improved LSTM neural network according to claim 1, characterized in that, The specific rules for dual-label coding in S5 are as follows: rockburst risk levels are divided into four levels, namely Level I (no rockburst), Level II (weak rockburst), Level III (medium rockburst), and Level IV (strong rockburst). The discrete labels use a combination of integer encoding and one-hot encoding. Level I corresponds to 0 and [1,0,0,0], Level II corresponds to 1 and [0,1,0,0], Level III corresponds to 2 and [0,0,1,0], and Level IV corresponds to 3 and [0,0,0,1]. Continuous labels linearly map discrete labels to continuous values ​​in the interval [0,1], with 0 corresponding to level I, 0.33 to level II, 0.66 to level III, and 1.0 to level IV; the label for each time series sample is based on the actual rockburst level corresponding to the last data point of that sample.

5. The rockburst risk classification and prediction method based on an improved LSTM neural network according to claim 2, characterized in that, The bidirectional interaction mechanism between the improved LSTM layer and the residual connection algorithm module in S3 is as follows: The residual connection algorithm module fuses and corrects the temporal features output by the first layer of the improved LSTM and the high-dimensional features output by the embedding layer to obtain the corrected features and inputs them into the second layer of the improved LSTM, providing a low-noise feature foundation for deep temporal feature extraction. During model training, the cross-entropy loss of the improved LSTM is backpropagated to the residual connection algorithm module to optimize its weight parameters, so that the feature correction can be adapted to the temporal feature extraction requirements of the improved LSTM.

6. The rockburst risk classification and prediction method based on an improved LSTM neural network according to claim 2, characterized in that, In S3, the SVR module interacts in a closed loop with the improved LSTM layer and the residual connection algorithm module. Specifically: the SVR module receives the corrected features output from the residual connection algorithm module and outputs continuously corrected values ​​to the attention layer, providing a quantitative basis for feature weight optimization; the regression loss of SVR and the cross-entropy loss of the improved LSTM work synergistically to update the weight parameters of the residual connection algorithm module; simultaneously, the fused features optimized by the attention layer are fed back to the SVR module, adjusting its kernel function parameters to adapt to the feature distribution; this closed-loop interaction continues until the total model loss decreases by less than [a certain value] for 10 consecutive training epochs. The iteration will end when the preset maximum number of iterations is reached.

7. The rockburst risk classification and prediction method based on an improved LSTM neural network according to claim 1, characterized in that, In S4, hyperparameter optimization adopts the grid search method. The hyperparameters to be optimized include the number of LSTM units, the dimension of the fully connected layer, the learning rate, the number of iterations, the batch size, the LSTM regularization coefficient, the penalty coefficient of SVR, the kernel function parameters, and the insensitive loss coefficient. The optimization objectives are the accuracy of the validation set and the mean of the F1 score. The early stopping strategy monitors and verifies the total loss or performance metrics. If the total loss does not decrease or the performance does not improve for 10 consecutive training cycles, training is stopped and the current best model is saved.

8. The rockburst risk classification and prediction method based on an improved LSTM neural network according to claim 1, characterized in that, In S6, the classification performance evaluation metrics include accuracy, precision, recall, and F1 score, while the regression performance evaluation metrics include mean squared error and coefficient of determination. The thresholds for achieving the desired results are: accuracy ≥ 85%, F1 score ≥ 80%, coefficient of determination ≥ 0.8, and total loss ≤ 0.

2. Once all thresholds are met, the model is determined as the final collaborative fusion prediction model. If the criteria are not met, the model is retrained and evaluated by adjusting the model parameters, supplementing time-series samples with the corresponding risk level, or optimizing the model architecture.

9. The rockburst risk classification and prediction method based on an improved LSTM neural network according to claim 1, characterized in that, The rockburst risk classification prediction results in S7 include core risk levels and auxiliary information. The core risk level is the level corresponding to the maximum value in the probability distribution. The auxiliary information includes the probability value of each risk level and the continuous correction value of SVR. If the maximum value in the probability distribution is <0.7, it is marked as a boundary risk sample, and the tendency level and corresponding enhanced monitoring recommendations are clearly stated.

10. The rockburst risk classification and prediction method based on an improved LSTM neural network according to claim 1, characterized in that, The data on factors influencing rockburst in S1 include: geostress parameters, rock mechanics parameters, engineering geological condition parameters, and construction disturbance parameters.

Images