Intelligent early warning method for deep underground engineering rockburst risk

Abstract

The invention provides a deep underground engineering rockburst risk intelligent early warning method, and constructs a set of rock mass dynamic disaster intelligent early warning system integrating micro-seismic signal frequency domain analysis, channel optimization and depth time sequence learning in order to solve the prevention and control problems of strong rockburst disaster burstiness, high spatial and temporal distribution randomness and the like in deep underground engineering. According to the method, high-timeliness and high-robustness prediction of rockburst key precursor parameters is realized, global extraction of microseismic signal frequency domain features is realized through an FFTConv1d module, transient and periodic components of microseismic signals are effectively captured, and feature completeness is improved; an SE attention mechanism self-adaptive weighted feature channel is introduced, the sensitivity to a microseismic signal critical precursor is enhanced, noise is suppressed, and the model robustness is improved; rockburst dynamic evolution of the cumulative apparent volume, the instantaneous apparent volume and the energy index is modeled in combination with xLSTM, and the core limitations of one-sided feature extraction, weak anti-noise ability, difficulty in capturing a nonlinear evolution rule, single early warning information and the like of a traditional method are overcome.

Description

Technical Field

[0001] This invention relates to the field of underground engineering safety monitoring technology, and in particular to an intelligent early warning method for rockburst risk in deep underground engineering. Background Technology

[0002] In deep underground engineering projects under high ground stress conditions, the construction and excavation strongly disturb the original rock stress field, which can easily induce sudden and destructive rockburst disasters, seriously threatening the safety of personnel and equipment and the progress of project construction. Accurate and real-time early warning of rockbursts in deep underground engineering projects is a key problem that urgently needs to be solved in the field of deep underground engineering safety.

[0003] Currently, engineering practice mainly relies on rockburst early warning technology systems centered on microseismic monitoring. However, existing methods have significant limitations in terms of timeliness, accuracy, and intelligence, specifically:

[0004] (1) Traditional methods are highly subjective and inefficient, making it difficult to meet the needs of real-time early warning. Existing technologies rely heavily on expert experience to manually interpret microseismic signals, requiring comprehensive analysis of time domain, frequency domain, and amplitude characteristics. This process is not only time-consuming and laborious, introducing significant time delays, but also suffers from poor consistency in interpretation results due to its high subjectivity, which seriously restricts the real-time performance and reliability of the early warning system.

[0005] (2) Traditional models are limited in their mechanisms and struggle to capture the nonlinear dynamic characteristics of the rockburst incubation process. Early warning methods based on empirical criteria or shallow statistical models, while simple and easy to implement, are insufficient to characterize the highly nonlinear and non-stationary temporal evolution of stress accumulation, crack initiation, and propagation during rockburst incubation. Numerical simulation methods, due to the applicability of constitutive relations, parameter uncertainties, and reliance on small deformation theory, have inherent limitations in reproducing the dynamic characteristics of rockburst suddenness and energy release. These methods are mostly suitable for long-term risk assessment and are ill-suited to addressing the dynamic evolution of risks during the construction period.

[0006] (3) Existing intelligent models have incomplete feature extraction and ignore key frequency domain precursor information. Although some inventions have attempted to introduce deep learning models such as long short-term memory networks and convolutional neural networks to improve prediction capabilities, most of these models focus on analysis from the time or spatial domain perspective. As a non-stationary time series, the frequency domain characteristics of microseismic signals are the key criteria for identifying rockburst precursors. Existing models generally lack efficient frequency domain feature extraction modules, resulting in insufficient utilization of the precursor information carried by frequency components, which limits further improvement in prediction accuracy.

[0007] (4) The models lack robustness and have a single prediction dimension. The construction site environment of deep underground engineering is complex, and microseismic signals are often mixed with strong interference such as blasting vibration and mechanical noise. Existing models generally lack targeted noise suppression and adaptive feature enhancement mechanisms, resulting in poor generalization ability and high false alarm rate under low signal-to-noise ratio conditions. In addition, most methods can only output the overall risk level and cannot achieve joint prediction and fine classification of multi-dimensional features of rockburst occurrence time, spatial location and intensity, making it difficult to support accurate prevention and control decisions that integrate "time-space-intensity".

[0008] In summary, existing technologies lack an intelligent early warning scheme that can efficiently integrate time-frequency domain features, adaptively suppress noise, and accurately predict the multi-dimensional characteristics and risk levels of rockbursts. This deficiency has become a technical bottleneck restricting the improvement of rockburst disaster prevention and control capabilities in deep underground engineering. Summary of the Invention

[0009] To address the problems existing in the prior art, this invention provides an intelligent early warning model and method that integrates frequency domain feature extraction, channel attention mechanism and extended long short-term memory network, aiming to achieve high-precision, real-time dynamic prediction and graded early warning of rockburst risk in multiple dimensions of time, space and intensity.

[0010] This invention provides an intelligent early warning method for rockburst risk in deep underground engineering, applicable to an intelligent early warning system for rockburst risk in deep underground engineering, including a microseismic sensor array, a data acquisition and transmission module, a central processing server, and an early warning terminal;

[0011] The microseismic sensor array includes several triaxial velocity-type microseismic sensors deployed in the working face of deep underground engineering and key areas of the surrounding rock to capture microseismic signals of rock mass fracture.

[0012] The data acquisition and transmission module includes a data acquisition instrument and a communication unit. The data acquisition instrument is connected to each micro-vibration sensor and is used to perform high-precision analog-to-digital conversion on the received analog electrical signals, converting them into discrete digital signals. The communication unit uses optical fiber communication technology to perform photoelectric conversion on the digital signals and then transmits them to the central processing server.

[0013] The central processing server deploys and runs a deep learning-based F-SE-xLSTM model to preprocess the received digital signals. The F-SE-xLSTM model is used to predict rockburst risk of the preprocessed digital signals. Based on the rockburst risk prediction results output by the model, the refined classification and confidence assessment of rockburst risk level are completed. The massive amount of microseismic event data, feature data and early warning results are structured, stored, managed and maintained to form a traceable rockburst incubation process database.

[0014] The early warning terminal communicates with the central processing server to visualize the prediction results of rockburst risk level, prediction confidence, and evolution trend of key microseismic parameters. It is also used to issue an alarm when the rockburst risk level exceeds the threshold, and to provide graded suggested prevention and control measures based on the rockburst risk level prediction results, providing clear decision support for proactive intervention.

[0015] The method includes:

[0016] S1. Collect historical microseismic signals from deep underground engineering, normalize them using the maximum absolute value scaling method, and preprocess them using a sliding time window mechanism to obtain a supervised learning sample set. Label the rockburst risk level of the supervised learning sample set, and divide the labeled supervised learning sample set into a training set, a validation set, and a test set for subsequent model training and performance evaluation.

[0017] S2. Construct and train the F-SE-xLSTM model, including the FFTConv1d module, the hierarchical feature fusion module, the improved SE module, and the early warning decision output module;

[0018] The FFTConv1d module is constructed to extract both the local time-domain features and the global frequency-domain features of the microseismic signal.

[0019] A hierarchical feature fusion module is constructed to fuse temporal local features and frequency domain global features through a multi-level attention mechanism;

[0020] An improved SE module is constructed to enhance key features and suppress noise through a channel-temporal dual attention mechanism;

[0021] Construct an early warning decision output module to predict rockburst risk levels;

[0022] During the training phase, the training set is input into the F-SE-xLSTM model to train the model's hyperparameters. During the validation phase, the validation set is input into the trained F-SE-xLSTM model to fine-tune the hyperparameters. During the testing phase, the test set is input into the constructed F-SE-xLSTM model to evaluate the model's prediction performance, thus obtaining the final F-SE-xLSTM model.

[0023] S3. Integrate new real-time microseismic signals from deep underground engineering projects, and use the maximum absolute value scaling method for normalization and the sliding time window mechanism for preprocessing the real-time microseismic signals.

[0024] S4. Input the preprocessed real-time microseismic signal into the trained F-SE-xLSTM model, and output the prediction results of the five-level rockburst risk level in the future time and the confidence level. When the rockburst risk level prediction results reach the preset threshold, multi-level early warning is automatically triggered.

[0025] Optionally, the construction of the FFTConv1d module simultaneously extracts the local time-domain features and global frequency-domain features of the microseismic signal, including:

[0026] An FFT convolution submodule is constructed to map the microseismic signal to the frequency domain using Fast Fourier Transform (FFT). Then, a one-dimensional convolution operation is performed in the frequency domain to extract key frequency domain features, including the dominant frequency, bandwidth, and spectral energy distribution, as detailed below:

[0027] ;

[0028] in, For Fast Fourier Transform, For inverse fast Fourier transform, For the original data points, For the filled input tensor, This is the Hadamard product (element-by-element multiplication). The convolution kernel after zero padding. These are bias parameters;

[0029] Introducing a frequency domain attention mechanism:

[0030] ;

[0031] ;

[0032] in, For attention weight tensors, It is the Sigmoid activation function. To correct the linear unit, The input feature tensor;

[0033] A multi-dimensional feature space that combines local temporal dynamics and global frequency response is constructed through temporal multi-scale convolution and frequency multi-scale convolution, as detailed below:

[0034] ;

[0035] ;

[0036] ;

[0037] .

[0038] Optionally, the hierarchical feature fusion module fuses temporal local features and frequency domain global features through a multi-level attention mechanism, including:

[0039] Initial feature fusion is achieved through linear transformation and layer normalization. Feature refinement is then performed by dynamically adjusting frequency weights and convolutional networks using a Sigmoid gating mechanism, as detailed below:

[0040] Through temporal self-attention:

[0041] ;

[0042] in, , , , The dimension of the key vector;

[0043] Frequency domain self-attention:

[0044] ;

[0045] Multi-layer fusion formula for the i-th layer:

[0046] ;

[0047] Cross-layer attention gating:

[0048] ;

[0049] ;

[0050] in, The weights are for the first-level linear transformation. For the weights of the second-level linear transformation, For the gated weight matrix, This is a characteristic of the historical layer.

[0051] Optionally, the improved SE module is constructed using a channel-temporal dual attention mechanism for key feature enhancement and noise suppression, including:

[0052] A dual attention mechanism of channel and time is constructed. Channel attention achieves feature recalibration through global average pooling and two fully connected layers, while time attention uses a one-dimensional convolutional structure to model temporal dependencies and enhance the discriminative ability of key features, as detailed below:

[0053] Channel attention (SE module):

[0054] ;

[0055] ;

[0056] ;

[0057] in, This is the global statistic for channel c. The length of the time series. Let be the eigenvalue of the c-th channel at time t. For channel attention weights, It is the ReLU activation function. The first layer of compression weights, For the second layer of expansion weights, The characteristics of the recalibrated channel;

[0058] An adaptively adjustable soft threshold structure is introduced into the SE module to adaptively shrink the feature map, enhance key feature channels related to rockburst precursors, and suppress non-key or interfering channels related to noise.

[0059] Time attention:

[0060] ;

[0061] ;

[0062] Depthwise separable convolution feature enhancement

[0063] ;

[0064] ;

[0065] .

[0066] Optionally, the early warning decision output module predicts the rockburst risk level based on the characteristics of refined microseismic signals, including:

[0067] The evolution of microseismic parameters is learned by using an extended long short-term memory network (xLSTM) to robustly extract rockburst precursor information and perform spatiotemporal joint prediction of risk status, outputting predicted values ​​of multiple key rockburst parameters in the future.

[0068] Based on the multi-parameter prediction results output by xLSTM, a fully connected classification layer is used to comprehensively determine and output the predicted probability of future rock eruptions. The prediction formula is as follows:

[0069] ;

[0070] in, Corresponding to five risk levels, For output layer weights, For classification bias vector, Risk level label;

[0071] Threshold determination for warning level:

[0072] .

[0073] By adopting the above technical solution, the present invention has at least the following beneficial effects:

[0074] This invention addresses the challenges of preventing and controlling rockburst disasters in deep underground engineering, characterized by their suddenness and highly random spatiotemporal distribution. It constructs an intelligent early warning system for rock mass dynamic disasters, centered on microseismic monitoring and incorporating multi-source information. Furthermore, it proposes a deep learning-based F-SE-xLSTM hybrid neural network model, achieving highly timely and robust prediction of key rockburst precursor parameters. The FFTConv1d module enables global extraction of frequency domain features from microseismic signals, effectively capturing transient and periodic components and improving feature completeness. An SE attention mechanism is introduced to adaptively weight the feature channels, enhancing sensitivity to critical precursors of microseismic signals, suppressing noise, and improving model robustness. Finally, by combining xLSTM modeling of the dynamic evolution of rockburst based on cumulative apparent volume, instantaneous apparent volume, and energy index, it overcomes the core limitations of traditional methods, such as one-sided feature extraction, weak noise resistance, difficulty in capturing nonlinear evolution patterns, and limited early warning information. This invention integrates microseismic signal frequency domain analysis, channel optimization, and deep time series learning, achieving a leap from a single indicator to a multi-dimensional risk assessment of "time-space-intensity" and supporting simultaneous prediction of multiple parameters. Attached Figure Description

[0075] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0076] Figure 1 This is a schematic diagram of an application scenario architecture upon which the present invention is based;

[0077] Figure 2 This is a schematic diagram of the F-SE-xLSTM model generation process proposed in this invention;

[0078] Figure 3 This is a schematic diagram of the structure of the F-SE-xLSTM model proposed in this invention;

[0079] Figure 4 Diagram of the adaptive soft threshold adjustment structure for the improved SE module;

[0080] Figure 5The following are the validation loss and accuracy curves for each model during the training process: (a) is the validation loss curve for each model during the training process; (b) is the validation accuracy curve for each model during the training process; and (c) is the microseismic energy prediction curve for each model during the training process.

[0081] Figure 6 This is a verification diagram for predicting an actual rockburst event in a tunnel.

[0082] Figure 7 The image shows the results of the confusion matrix analysis for the F-SE-xLSTM model. Detailed Implementation

[0083] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0084] The purpose of this invention is to overcome the inherent defects of existing rockburst early warning technologies for deep underground engineering, such as delayed early warning, high false alarm rate, incomplete feature extraction, and single prediction dimension, and to provide a high-precision, high-efficiency, and highly robust rockburst risk prediction method.

[0085] To achieve the above-mentioned objectives, this invention provides an intelligent early warning method for rockburst risk in deep underground engineering based on multi-domain feature fusion. The core of this method is to construct and apply a multi-domain fusion deep learning model called F-SE-xLSTM. This model mainly includes an FFTConv1d module, a hierarchical feature fusion module, an improved SE module, and an early warning decision output module, which systematically solves the end-to-end technical challenges from feature extraction to rockburst risk output.

[0086] Figure 1 This is a schematic diagram of an application scenario architecture upon which the present invention is based. It is implemented by an intelligent early warning system for rockburst risk in deep underground engineering, including a microseismic sensor array, a data acquisition and transmission module, a central processing server, and an early warning terminal.

[0087] 1. Microseismic sensor array

[0088] The microseismic sensor array is the front-end sensing unit of the system, consisting of multiple high-sensitivity, triaxial velocity-type microseismic sensors. The sensors are preferably of the type with a natural frequency of 14Hz and a response frequency band of 10–1000Hz to match the dominant frequency range generated by the micro-fractures in the rock mass. The sensors are deployed according to the principle of spatial optimization, focusing on covering the potential high-risk areas 30 meters in front of and 20 meters behind the working face of deep underground engineering projects, forming a three-dimensional monitoring network for capturing elastic wave (P-wave and S-wave) signals generated by micro-fractures in the surrounding rock in an all-round and real-time manner.

[0089] 2. Data Acquisition and Transmission Module

[0090] The data acquisition and transmission module is the system's "nerve" conduction unit, responsible for the preliminary processing and reliable transmission of microseismic signals, including:

[0091] Data acquisition unit: Connected to each micro-vibration sensor, it is used to perform high-precision analog-to-digital conversion on the received analog electrical signals, converting them into discrete digital signals.

[0092] Communication Unit: Utilizes fiber optic communication technology to convert digital signals into photoelectric signals and transmit them to the central processing server, enabling remote, low-loss, high-fidelity, and interference-resistant data transmission, ensuring that signals can be delivered to the central server completely and in real time.

[0093] 3. Central Processing Server

[0094] The central processing server is the "brain" and computing core of the system, deploying and running the deep learning-based F-SE-xLSTM model. The central processing server undertakes the following core tasks:

[0095] The received digital signals are preprocessed, and the F-SE-xLSTM model is used to predict the rockburst risk of the preprocessed digital signals. Based on the rockburst risk prediction results output by the model, the rockburst risk level is refined and the confidence level is assessed. The massive amount of microseismic event data, feature data and early warning results are structured, stored, managed and maintained to form a traceable rockburst incubation process database.

[0096] 4. Early warning terminal

[0097] The early warning terminal is the system's human-computer interaction and decision support unit. It communicates with the central processing server and has the following functions:

[0098] Risk visualization: Visualize the prediction results of rockburst risk level, prediction confidence, and the evolution trend of key microseismic parameters.

[0099] Alarms and prompts: When the rockburst risk level exceeds the threshold, an alarm will be triggered. At the same time, based on the rockburst risk level prediction results, suggested prevention and control measures will be provided at different levels to provide clear decision support for proactive intervention.

[0100] The following will further explain the intelligent early warning method for rockburst risk in deep underground engineering provided in this application:

[0101] This disclosure provides an intelligent early warning method for rockburst risk in deep underground engineering, applicable to the aforementioned intelligent early warning system for rockburst risk in deep underground engineering, including:

[0102] 1. System Deployment and Data Acquisition

[0103] In a high-risk rockburst section of a deep-buried tunnel on a plateau, based on an optimized topology design, a three-dimensional velocity-type microseismic sensor array was deployed within a range of 30 meters in front of the tunnel face and 20 meters behind the tunnel face. The preferred sensor model has a natural frequency of 14Hz and a response frequency band of 10–1000Hz to accurately capture rock fracture signals. The sampling frequency of the data acquisition instrument was set to 4kHz, and a central processing server was deployed at the control center outside the tunnel.

[0104] The system uses an array of microseismic sensors deployed at the tunnel face and key areas of the surrounding rock to collect 30,000 sets of historical microseismic signals from a deep-buried tunnel on a plateau and transmits them in real time to the central processing server at the control center outside the tunnel.

[0105] The process of using collected historical microseismic signals to construct a dataset and train a rockburst risk prediction model is as follows: Figure 2 As shown, this corresponds to steps 2-4.

[0106] 2. Data Preprocessing and Dataset Construction

[0107] The historical microseismic signals were normalized using the maximum absolute value scaling method to eliminate amplitude deviations caused by sensor sensitivity or transmission distance, as detailed below:

[0108] ;

[0109] ;

[0110] in, , For the original data points, For batch size, The sequence length dimension For the feature vector dimension, , For the sample mean and standard deviation, It is the numerical stability constant. , These are learnable parameters.

[0111] A sliding time window mechanism is used to preprocess historical microseismic signals to smooth short-term fluctuations, reduce computational overhead, and enhance the model's generalization ability, resulting in a supervised learning sample set, as follows:

[0112] .

[0113] T is the original sequence length, W is the target window length, where ":" indicates Python slicing syntax, and the first ":" means to retain all batch samples. " is the starting index, and the last ":" means to retain all feature dimensions.

[0114] in, To standardize the processed data tensor, based on previous parameter analysis, the optimal window length was determined to be 150 hours. The window sliding step size can be set according to the actual warning refresh rate requirements (1 hour). Each sample contains data of 12-dimensional microseismic parameters (such as the number of microseismic events, energy, apparent volume, energy index, etc.) over 128 consecutive time steps.

[0115] The supervised learning sample set is labeled with rockburst risk level, and the labeled supervised learning sample set is divided into training set, validation set and test set for subsequent model training and performance evaluation.

[0116] 3. Construct and train the F-SE-xLSTM model

[0117] like Figure 3 As shown, the constructed F-SE-xLSTM model includes an FFTConv1d module, a hierarchical feature fusion module, an improved SE module, and an early warning decision output module.

[0118] (1) FFTConv1d module

[0119] Function: Simultaneously extracts local time-domain features and global frequency-domain features of microseismic signals.

[0120] An FFT convolution submodule is constructed to map the microseismic signal to the frequency domain using Fast Fourier Transform (FFT). Then, a one-dimensional convolution operation is performed in the frequency domain to extract key frequency domain features, including the dominant frequency, bandwidth, and spectral energy distribution, as detailed below:

[0121] ;

[0122] in, For Fast Fourier Transform, For inverse fast Fourier transform, For the original data points, For the filled input tensor, This is the Hadamard product (element-by-element multiplication). The convolution kernel after zero padding. This is the bias parameter.

[0123] Introducing a frequency domain attention mechanism (global average pooling → 1 / 4 channel dimensionality reduction → ReLU activation → channel restoration) significantly enhances feature discrimination capabilities:

[0124] ;

[0125] ;

[0126] in, For attention weight tensors, It is the Sigmoid activation function. To correct the linear unit, The input feature tensor.

[0127] By constructing a multi-dimensional feature space that combines local temporal dynamics and global frequency response through temporal multi-scale convolution and frequency multi-scale convolution, the completeness of feature representation is significantly improved, as detailed below:

[0128] ;

[0129] ;

[0130] ;

[0131] .

[0132] Compared to conventional temporal convolutional layers, the FFTConv1d module has the following advantages:

[0133] 1) The receptive field of traditional temporal convolution is limited by the kernel size, making it difficult to effectively capture long-range dependencies. However, the FFTConv1d module performs frequency domain transformation based on the entire signal sequence, which can extract the frequency response and periodic patterns across the time sequence from a global perspective in one go, significantly enhancing the ability to identify the frequency domain features of rockburst precursors.

[0134] 2) As a non-stationary time series, the frequency domain characteristics of microseismic signals are a key criterion for identifying rockburst risk. The FFTConv1d module performs convolution operations and feature learning directly in the frequency domain, avoiding the computational redundancy and learning inefficiency caused by the implicit extraction of frequency domain features through deep networks in traditional time-domain convolution models. This enables the model to more accurately and directly capture the frequency characteristics closely related to the rockburst mechanism and their time-varying evolution.

[0135] 3) Microseismic signals collected in actual construction environments are often accompanied by a large amount of environmental noise and equipment interference. Since noise is mostly concentrated in specific frequency bands, frequency domain analysis is more conducive to distinguishing rockburst signals from noise components. The FFTConv1d module effectively suppresses interference from irrelevant frequency bands by introducing an adaptive filtering mechanism in the frequency domain, enhances the representation strength of effective microseismic signals, and thus improves the quality of input features and the generalization performance of the model in complex engineering scenarios.

[0136] (2) Layered feature fusion module

[0137] Function: It fuses temporal local features and frequency domain global features through a multi-level attention mechanism.

[0138] Initial feature fusion is achieved through linear transformation and layer normalization. Feature refinement is then performed by dynamically adjusting frequency weights and convolutional networks using a Sigmoid gating mechanism, as detailed below:

[0139] Through temporal self-attention:

[0140] ;

[0141] in, , , , The dimension of the key vector.

[0142] Frequency domain self-attention:

[0143] ;

[0144] Multi-layer fusion formula for the i-th layer:

[0145] ;

[0146] Cross-layer attention gating:

[0147] ;

[0148] ;

[0149] in, The weights are for the first-level linear transformation. For the weights of the second-level linear transformation, For the gated weight matrix, This is a characteristic of the historical layer.

[0150] (3) Improved SE module

[0151] Function: Enhances key features and suppresses noise through a channel-temporal dual attention mechanism.

[0152] A dual attention mechanism of channel and time is constructed. Channel attention achieves feature recalibration through global average pooling and two fully connected layers, while time attention uses a one-dimensional convolutional structure to model temporal dependencies and enhance the discriminative ability of key features, as detailed below:

[0153] Channel attention (SE module):

[0154] ;

[0155] ;

[0156] ;

[0157] in, This is the global statistic for channel c. The length of the time series. Let be the eigenvalue of the c-th channel at time t. For channel attention weights, It is the ReLU activation function. The first layer of compression weights, For the second layer of expansion weights, The channel characteristics after recalibration.

[0158] Traditional SENet lacks explicit noise suppression mechanisms, while microseismic signals in practical engineering are often affected by strong environmental noise and equipment interference. To address this challenge, such as Figure 4 As shown, an adaptively adjustable soft threshold structure is introduced into the SE module to adaptively shrink the feature map, strengthen the key feature channels related to rockburst precursors, and suppress non-key or interfering channels related to noise, effectively improving the robustness of the model in high-noise environments.

[0159] Time attention:

[0160] ;

[0161] ;

[0162] Depthwise separable convolution feature enhancement

[0163] ;

[0164] ;

[0165] .

[0166] (4) Early warning decision output module

[0167] Function: Predicts rockburst risk level.

[0168] The evolution of microseismic parameters is learned by using an extended long short-term memory network (xLSTM), and a variant of mLSTM is used as a temporal decoder. By constructing matrix memory units and a covariance update mechanism based on key-value pair storage, it overcomes the limitations of traditional LSTM in long-term dependency modeling and memory capacity, and performs robust extraction of rockburst precursor information and spatiotemporal joint prediction of risk status. It outputs predicted values ​​of multiple key rockburst parameters (including but not limited to cumulative energy E, seismic moment M0, cumulative apparent volume VA, and energy index EI) for the future.

[0169] Based on the multi-parameter prediction results output by xLSTM, a fully connected classification layer is used to comprehensively determine and output the predicted probability of future rock eruptions. The prediction formula is as follows:

[0170] ;

[0171] in, Corresponding to five risk levels, For output layer weights, This is the output layer bias vector. This is a risk level label.

[0172] Threshold determination for warning level:

[0173] The constructed F-SE-xLSTM model is trained, and the key training parameters are as follows:

[0174] Input dimension: 12 (feature dimension) × 128 (time steps).

[0175] Loss function: For regression prediction tasks involving parameters such as cumulative apparent volume and energy index, mean squared error (MSE) is used as the loss function; for classification tasks with five risk levels, cross-entropy loss function is used.

[0176] Optimizer and Hyperparameters: The Adam optimizer was used with an initial learning rate of 0.005; the batch size was set to 64; Dropout was introduced in parameter-dense layers such as fully connected layers with a dropout rate of 0.2 to prevent overfitting; Batch Normalization (BatchNorm) layers were used to accelerate convergence with a momentum parameter of 0.1. Model training continued until the loss function on the validation set no longer decreased significantly, and the optimal model weights were saved.

[0177] During the testing phase, the test set is input into the constructed F-SE-xLSTM model to evaluate the model's prediction performance and obtain the final F-SE-xLSTM model.

[0178] 4. Performance Evaluation and Comparison

[0179] To verify the effectiveness and accuracy of the F-SE-xLSTM model, three mature models—CNN-FNN, LSTM-FCNN, and HCLSTM—were selected as comparison models for performance evaluation and comparison. To control for differences in model complexity, all comparison models adopted the standard configuration with the fewest parameters and were tested in a unified training environment. To ensure fairness in the comparison, the hyperparameter settings of each model were consistent with those of the F-SE-xLSTM model. For key parameters such as the learning rate, a grid search method was used for fine-tuning with validation set performance as the optimization objective. Figure 5 The validation loss versus accuracy curves for each model during training are shown. For example... Figure 5 As shown in (a), the losses of all models decrease rapidly in the initial stage, but the F-SE-xLSTM model exhibits the best convergence characteristics, stabilizing after approximately 40 training epochs, with a final validation loss of 0.019 and a validation accuracy of 95.5%. This rapid convergence capability indicates that the SE module significantly improves the model's perception efficiency of rockburst precursor time-series features through feature recalibration, which is beneficial for rapid deployment and iterative updates of the model in practical engineering scenarios. Figure 5 As shown in (b), the accuracy curves exhibit a trend highly consistent with the validation loss. The F-SE-xLSTM model rapidly improves its accuracy from 75% to 90% within the first 30 rounds, subsequently stabilizing at 95.5%, indicating that its architecture can efficiently capture typical dynamic patterns during rockburst incubation. In summary, SE-xLSTM, with its high prediction accuracy and rapid convergence, can still achieve reliable early warning under limited computational resources, making it suitable for engineering applications requiring real-time rockburst monitoring systems.

[0180] To visually compare the performance of each model in rockburst prediction tasks, microseismic energy was selected as the key rockburst indicator output by the models. Figure 5 (c) shows a comparison of the prediction results of the four models. The results show that there are significant differences in prediction accuracy and dynamic response characteristics among the different models: the F-SE-xLSTM model exhibits the best overall performance, not only with the highest prediction accuracy, but also with the fastest response to rockburst precursor signals.

[0181] The mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), coefficient of determination (R²), and mean absolute percentage error (MAPE) are selected as the model accuracy evaluation indicators. As can be seen from Table 1, the F-SE-xLSTM model performs excellently in all indicators: the RMSE is 3.1, which is significantly reduced by 85.8% compared with the CNN-FNN model (RMSE = 21.8); the MAE is 2.2, which is improved by 80.5% compared with the CNN-FNN model (MAE = 11.3); the R² reaches 0.99, indicating a high degree of consistency between the predicted values and the true values. These indicators are sufficient to show that the F-SE-xLSTM model proposed in this invention can effectively capture the temporal evolution characteristics of microseismic parameters and provide reliable technical support for rockburst monitoring and early warning.

[0182] Table 1 Precision evaluation indicators of four models

[0183]

[0184] 5. Model Deployment and Online Early Warning

[0185] The trained F-SE-xLSTM model is deployed to the microseismic monitoring system software platform of the central processing server, new real-time microseismic signals of the deep-buried tunnel are accessed, and normalization is carried out using the maximum absolute value scaling method and the real-time microseismic signals are preprocessed using the sliding time window mechanism. The preprocessed real-time microseismic signals are input into the trained F-SE-xLSTM model, and the prediction results of the five-level rockburst risk level and the confidence level for future time are output. When the prediction result of the rockburst risk level reaches the preset threshold, multi-level early warning is automatically triggered.

[0186] Predictions are executed once every 1 hour, and the evolution curves of key rockburst parameters (cumulative energy, cumulative apparent volume, energy index) within the next 6 hours and the corresponding five-level risk levels are output. When the risk level predicted by the model reaches or exceeds level IV (high risk) and the prediction confidence level is higher than 90%, the system automatically triggers the following linkage response, activates the audible and visual alarms in the tunnel and the command center, and pushes high-level early warning information to the mobile terminals and management platforms of the construction responsible person and safety engineer, including the risk level, estimated time period, location of the high-risk area, and recommended measures.

[0187] In practical application, the model successfully provided an early warning for a moderate-intensity rockburst. On June 15, 2025, a moderate-intensity transient rockburst occurred at the tunnel face at DK1031+665.2. Based on data from the previous 7 days, the model accurately predicted the key precursor trends of the increase in cumulative apparent volume and the decrease in energy index, providing a response window of approximately 1.5–2.0 hours for the site. This effectively supported the implementation of proactive prevention and control measures, verifying the high accuracy and engineering practicality of the method of this invention. The event was fully recorded by the microseismic monitoring system. Figure 6 This study presents the temporal evolution characteristics and model prediction results of microseismic activity over three days before and after a rockburst (June 14-16, 2025). Monitoring data shows that during the rockburst incubation stage, both the cumulative apparent volume and energy index remained stable. However, at 11:27 AM on June 15th, the moment of the rockburst, both the cumulative and instantaneous apparent volumes showed a significant surge, while the energy index exhibited a rapid decline. A prediction model built based on microseismic time-frequency waveforms and parameter data from the seven days prior to the rockburst (June 7-13) successfully predicted the evolution trends of key parameters such as cumulative apparent volume, instantaneous apparent volume, and energy index. Although there were some discrepancies between the predicted and measured values, the model accurately captured the core characteristics of rockburst precursors—the significant increase in cumulative and instantaneous apparent volumes, and the sharp decline in the energy index. This parameter evolution model reveals the dynamic accumulation mechanism of rockburst risk: under high geostress conditions, rock mass stress continuously accumulates and gradually releases energy through microseismic activity; when the accumulated energy exceeds the rock mass strength threshold, sudden instability is triggered, leading to a rockburst. Following a rockburst event, the elastic strain energy stored within the rock mass is rapidly released, and the level of microseismic activity subsequently stabilizes. This case study validates the effectiveness of the proposed model in identifying precursory risks of rockbursts, providing a crucial time window for the implementation of proactive prevention and control measures such as support reinforcement and personnel evacuation.

[0188] Based on data from 5000 rockburst engineering cases, a five-level rockburst hazard classification and prediction model was constructed, corresponding to risk levels as: Level I (no risk), Level II (low risk), Level III (medium risk), Level IV (high risk), and Level V (extremely high risk). Figure 7The confusion matrix analysis results show that the model exhibits excellent classification performance across all risk levels. Overall, the model achieves precision and recall exceeding 95% on the test set, validating the high reliability of its predictions. The classification performance is most outstanding for Level III (medium risk), with both precision and recall reaching 96.8%. Error distribution analysis reveals that misclassifications are mainly concentrated between adjacent risk levels, with the highest proportion (8.0%) of Level IV being misclassified as Level V. This misclassification pattern aligns with the engineering reality of continuous evolution of rockburst risk, while also reflecting reasonable uncertainty in the model's identification of risk level boundary regions. The rockburst prediction model proposed in this invention significantly improves the ability to identify high-risk levels while maintaining high overall accuracy. The recall rates for Level IV and Level V risks reach 95.8% and 95.5%, respectively, indicating that the model can effectively capture potential high-risk conditions and provide scientific decision support for tunnel construction safety risk management.

[0189] The present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A method for intelligent early warning of rockburst risk in deep underground engineering, applicable to an intelligent early warning system for rockburst risk in deep underground engineering, comprising a microseismic sensor array, a data acquisition and transmission module, a central processing server, and an early warning terminal; The microseismic sensor array includes several triaxial velocity-type microseismic sensors deployed in the working face of deep underground engineering and key areas of the surrounding rock to capture microseismic signals of rock mass fracture. The data acquisition and transmission module includes a data acquisition instrument and a communication unit. The data acquisition instrument is connected to each micro-vibration sensor and is used to perform high-precision analog-to-digital conversion on the received analog electrical signals, converting them into discrete digital signals. The communication unit uses optical fiber communication technology to perform photoelectric conversion on the digital signals and then transmits them to the central processing server. The central processing server deploys and runs a deep learning-based F-SE-xLSTM model to preprocess the received digital signals. The F-SE-xLSTM model is used to predict rockburst risk of the preprocessed digital signals. Based on the rockburst risk prediction results output by the model, the refined classification and confidence assessment of rockburst risk level are completed. The massive amount of microseismic event data, feature data and early warning results are structured, stored, managed and maintained to form a traceable rockburst incubation process database. The early warning terminal communicates with the central processing server to visualize the prediction results of rockburst risk level, prediction confidence, and evolution trend of key microseismic parameters. It is also used to issue an alarm when the rockburst risk level exceeds the threshold, and to provide graded suggested prevention and control measures based on the rockburst risk level prediction results, providing clear decision support for proactive intervention. Its features are, The method includes: S1. Collect historical microseismic signals from deep underground engineering, normalize them using the maximum absolute value scaling method, and preprocess them using a sliding time window mechanism to obtain a supervised learning sample set. Label the rockburst risk level of the supervised learning sample set, and divide the labeled supervised learning sample set into a training set, a validation set, and a test set for subsequent model training and performance evaluation. S2. Construct and train the F-SE-xLSTM model, including the FFTConv1d module, the hierarchical feature fusion module, the improved SE module, and the early warning decision output module; The FFTConv1d module is constructed to extract both the local time-domain features and the global frequency-domain features of the microseismic signal. A hierarchical feature fusion module is constructed to fuse temporal local features and frequency domain global features through a multi-level attention mechanism; An improved SE module is constructed to enhance key features and suppress noise through a channel-temporal dual attention mechanism; Construct an early warning decision output module to predict rockburst risk levels; During the training phase, the training set is input into the F-SE-xLSTM model to train the model's hyperparameters. During the validation phase, the validation set is input into the trained F-SE-xLSTM model to fine-tune the hyperparameters. During the testing phase, the test set is input into the constructed F-SE-xLSTM model to evaluate the model's prediction performance, thus obtaining the final F-SE-xLSTM model. S3. Integrate new real-time microseismic signals from deep underground engineering projects, and use the maximum absolute value scaling method for normalization and the sliding time window mechanism for preprocessing the real-time microseismic signals. S4. Input the preprocessed real-time microseismic signal into the trained F-SE-xLSTM model, and output the prediction results of the five-level rockburst risk level in the future time and the confidence level. When the rockburst risk level prediction results reach the preset threshold, multi-level early warning is automatically triggered.

2. The intelligent early warning method for rockburst risk in deep underground engineering according to claim 1, characterized in that, The constructed FFTConv1d module simultaneously extracts the local time-domain features and global frequency-domain features of the microseismic signal, including: An FFT convolution submodule is constructed to map the microseismic signal to the frequency domain using Fast Fourier Transform (FFT). Then, a one-dimensional convolution operation is performed in the frequency domain to extract key frequency domain features, including the dominant frequency, bandwidth, and spectral energy distribution, as detailed below: ； in, For Fast Fourier Transform, For inverse fast Fourier transform, For the original data points, For the filled input tensor, This is the Hadamard product (element-by-element multiplication). The convolution kernel after zero padding. These are bias parameters; Introducing a frequency domain attention mechanism: ； ； in, For attention weight tensors, It is the Sigmoid activation function. To correct the linear unit, The input feature tensor; A multi-dimensional feature space that combines local temporal dynamics and global frequency response is constructed through temporal multi-scale convolution and frequency multi-scale convolution, as detailed below: ； ； ； 。 3. The intelligent early warning method for rockburst risk in deep underground engineering according to claim 1, characterized in that, The hierarchical feature fusion module fuses temporal local features and frequency domain global features through a multi-level attention mechanism, including: Initial feature fusion is achieved through linear transformation and layer normalization. Feature refinement is then performed by dynamically adjusting the frequency weights and convolutional networks using the Sigmoid gating mechanism, as detailed below: Through temporal self-attention: ； in, , , , The dimension of the key vector; Frequency domain self-attention: ； Multi-layer fusion formula for the i-th layer: ； Cross-layer attention gating: ； ； in, The weights are for the first-level linear transformation. For the weights of the second-level linear transformation, For the gated weight matrix, This is a characteristic of the historical layer.

4. The intelligent early warning method for rockburst risk in deep underground engineering according to claim 1, characterized in that, The improved SE module is constructed using a channel-temporal dual attention mechanism for key feature enhancement and noise suppression, including: A dual attention mechanism of channel and time is constructed. Channel attention achieves feature recalibration through global average pooling and two fully connected layers, while time attention uses a one-dimensional convolutional structure to model temporal dependencies and enhance the discriminative ability of key features, as detailed below: Channel attention (SE module): ； ； ； in, This is the global statistic for channel c. The length of the time series. Let be the eigenvalue of the c-th channel at time t. For channel attention weights, It is the ReLU activation function. The first layer of compression weights, For the second layer of expansion weights, The characteristics of the recalibrated channel; An adaptively adjustable soft threshold structure is introduced into the SE module to adaptively shrink the feature map, enhance key feature channels related to rockburst precursors, and suppress non-key or interfering channels related to noise. Time attention: ； ； Depthwise separable convolution feature enhancement ； ； 。 5. The intelligent early warning method for rockburst risk in deep underground engineering according to claim 1, characterized in that, The early warning decision output module predicts the rockburst risk level based on the characteristics of refined microseismic signals, including: The evolution of microseismic parameters is learned by using an extended long short-term memory network (xLSTM) to robustly extract rockburst precursor information and perform spatiotemporal joint prediction of risk status, outputting predicted values ​​of multiple key rockburst parameters in the future. Based on the multi-parameter prediction results output by xLSTM, a fully connected classification layer is used to comprehensively determine and output the predicted probability of future rock eruptions. The prediction formula is as follows: ； in, Corresponding to five risk levels, For output layer weights, For classification bias vector, Risk level label; Threshold determination for warning level: 。

Images