CN-121978757-A - Intelligent early warning method for rock burst risk of deep underground engineering

CN121978757ACN 121978757 ACN121978757 ACN 121978757ACN-121978757-A

Abstract

The invention provides an intelligent rock burst risk early warning method for deep underground engineering, which aims at the problems of strong rock burst disaster burst, high space-time distribution randomness and the like in the deep underground engineering, builds a set of intelligent rock mass dynamic disaster early warning system integrating micro-seismic signal frequency domain analysis, channel optimization and deep time sequence learning, realizes the high timeliness and high robustness prediction of key precursor parameters of rock burst, realizes the global extraction of the frequency domain characteristics of the micro-seismic signal through a FFTConv d module, effectively captures transient and periodic components of the micro-seismic signal, improves the characteristic completeness, introduces an SE attention mechanism self-adaptive weighted characteristic channel, enhances the sensitivity to the critical precursor of the micro-seismic signal, suppresses noise, improves model robustness, and combines xLSTM modeling to accumulate the rock burst dynamic evolution of a visual volume, an instantaneous visual volume and an energy index, thereby overcoming the core limitations of the traditional method such as one-sided feature extraction, weak anti-noise capability, difficulty in capturing nonlinear evolution law, single early warning information and the like.

Inventors

DAI FENG
DA YUXIN
LIU YI
WEI MINGDONG
YAN ZELIN

Assignees

四川大学

Dates

Publication Date: 20260505
Application Date: 20260121

Claims (5)

1. The intelligent pre-warning method for the rock burst risk of the deep underground engineering is suitable for an intelligent pre-warning system for the rock burst risk of the deep underground engineering, and comprises a microseismic sensor array, a data acquisition and transmission module, a central processing server and a pre-warning terminal; the microseismic sensor array comprises a plurality of three-way speed microseismic sensors which are distributed on a deep underground engineering working surface and a surrounding rock key area and are used for capturing rock mass fracture microseismic signals; the data acquisition and transmission module comprises a data acquisition instrument and a communication unit, wherein the data acquisition instrument is connected with each microseismic sensor and is used for carrying out high-precision analog-to-digital conversion on the received analog electric signals, converting the analog electric signals into discrete digital signals, and transmitting the digital signals to the central processing server after photoelectric conversion by the communication unit by adopting an optical fiber communication technology; The central processing server deploys and runs an F-SE-xLSTM model based on deep learning, is used for preprocessing a received digital signal, carries out rock burst risk prediction on the preprocessed digital signal by using the F-SE-xLSTM model, finishes fine classification and confidence assessment of rock burst risk level based on a rock burst risk prediction result output by the model, and carries out structural storage, management and maintenance on massive microseismic event data, characteristic data and early warning results to form a traceable rock burst inoculation process database; The early warning terminal is in communication connection with the central processing server, and is used for visually displaying a rock burst risk level prediction result, prediction confidence coefficient and an evolution trend of a key microseismic parameter, alarming when the rock burst risk level exceeds a threshold value, providing a graded suggestion prevention and control measure according to the rock burst risk level prediction result and providing clear decision support for active intervention; Characterized in that the method comprises: S1, collecting historical microseismic signals of deep underground engineering, carrying out normalization processing by adopting a maximum absolute value scaling method, preprocessing the historical microseismic signals by adopting a sliding time window mechanism to obtain a supervision learning sample set, carrying out rock burst risk grade marking on the supervision learning sample set, and dividing the marked supervision learning sample set into a training set, a verification set and a test set for subsequent model training and performance evaluation; S2, constructing and training an F-SE-xLSTM model, wherein the F-SE-xLSTM model comprises a FFTConv d module, a layered feature fusion module, an improved SE module and an early warning decision output module; constructing FFTConv d module to extract local time domain feature and global frequency domain feature of microseismic signal; A hierarchical feature fusion module is constructed to fuse the local time domain features and the global frequency domain features through a multi-level attention mechanism; Constructing an improved SE module, and carrying out key feature enhancement and noise suppression through a channel-time dual attention mechanism; Constructing an early warning decision output module to predict the rock burst risk level; Inputting a training set into the F-SE-xLSTM model to train the model super-parameters in a training stage, inputting a verification set into the trained F-SE-xLSTM model to fine tune the super-parameters in a verification stage, inputting a test set into the constructed F-SE-xLSTM model to evaluate the model prediction performance in a test stage, and obtaining a final F-SE-xLSTM model; s3, accessing a new real-time microseismic signal of the deep underground engineering, carrying out normalization processing by adopting a maximum absolute value scaling method, and carrying out pretreatment on the real-time microseismic signal by adopting a sliding time window mechanism; s4, inputting the preprocessed real-time microseismic signals into a trained F-SE-xLSTM model, outputting a five-level rock burst risk level prediction result and confidence level in future time, and automatically triggering multi-level early warning when the rock burst risk level prediction result reaches a preset threshold value.
2. The intelligent pre-warning method for the rock burst risk of the deep underground engineering according to claim 1, wherein the constructing FFTConv d module extracts the local time domain feature and the global frequency domain feature of the microseismic signal at the same time, and the method comprises the following steps: An FFT convolution sub-module is constructed, a microseismic signal is mapped to a frequency domain through Fast Fourier Transform (FFT), then one-dimensional convolution operation is carried out in the frequency domain, and key frequency domain characteristics including main frequency, bandwidth and spectrum energy distribution are extracted, wherein the key frequency domain characteristics are specifically as follows: ; Wherein, the In the case of a fast fourier transform, In the form of an inverse fast fourier transform, As the raw data point of the data, In order to fill in the input tensor, For hadamard product (element-wise multiplication), Is the convolution kernel after zero-padding, Is a bias parameter; introducing a frequency domain attention mechanism: ; ; Wherein, the In order to pay attention to the weight tensor, The function is activated for Sigmoid, In order to correct the linear units, Is an input feature tensor; the multidimensional feature space with local time domain dynamic and global frequency domain response is constructed through time domain multiscale convolution and frequency domain multiscale convolution, and the method is specifically as follows: ; ; ; 。
3. the intelligent pre-warning method for the rock burst risk of the deep underground engineering according to claim 1, wherein the construction layering feature fusion module fuses the local time domain feature and the global frequency domain feature through a multi-level attention mechanism, and the method comprises the following steps: Preliminary feature fusion is realized through linear transformation and layer normalization, and frequency weights and a convolution network are dynamically adjusted by using a Sigmoid gating mechanism to refine features, wherein the method comprises the following steps of: By time domain self-attention: ; Wherein, the , , , Is a key vector dimension; frequency domain self-attention: ; Multilayer fusion the i-th layer fusion formula: ; Cross-layer attention gating: ; ; Wherein, the For the first layer of linear transform weights, For the second layer of linear transformation weights, In order to gate the weight matrix, Is a history layer feature.
4. The intelligent pre-warning method for rock burst risk of deep underground engineering according to claim 1, wherein the construction of the improved SE module performs key feature enhancement and noise suppression through a channel-time dual attention mechanism, comprising: Constructing a channel-time dual-attention mechanism, realizing feature recalibration by the channel attention through global average pooling and double-layer full-connection layers, and enhancing the discrimination capability of key features by using a one-dimensional convolution structure modeling time sequence dependency relationship by the time attention, wherein the method comprises the following specific steps: Channel attention (SE module): ; ; ; Wherein, the As a global statistic of the channel c, For the length of the time series, For the characteristic value of the c-th channel at time t, For the channel attention weighting to be used, For the function to be activated by the ReLU, For the first layer of the laminated weights, The weights are expanded for the second layer, Channel characteristics after recalibration; introducing a soft threshold structure capable of being adaptively adjusted into the SE module, adaptively performing contraction operation on the feature map, strengthening key feature channels related to rock burst precursors, and inhibiting non-key or interference channels related to noise; Time attention: ; ; Depth separable convolution feature enhancement ; ; 。
5. The intelligent pre-warning method for rock burst risk of deep underground engineering according to claim 1, wherein the construction pre-warning decision output module predicts the rock burst risk level according to the characteristics of the refined microseismic signals, and comprises the following steps: Adopting an extended long-short-term memory network xLSTM to learn the evolution rule of microseismic parameters, carrying out robust extraction of rock burst precursor information and space-time combined prediction of risk states, and outputting predicted values of a plurality of key rock burst parameters at future time; based on xLSTM output multiparameter prediction results, comprehensively judging and outputting prediction results of the rock burst occurrence probability at future time through a fully connected classification layer, wherein the prediction formula is as follows: ; Wherein, the Corresponding to the five risk levels, In order to output the layer weights, In order to classify the offset vector, Is a risk level label; Threshold judgment is carried out on the early warning level: 。

Description

Intelligent early warning method for rock burst risk of deep underground engineering Technical Field The invention relates to the technical field of underground engineering safety monitoring, in particular to an intelligent early warning method for rock burst risk of deep underground engineering. Background In deep underground engineering under high ground stress conditions, construction excavation severely disturbs a primary rock stress field, rock explosion disasters with strong burst property and high destructiveness are extremely easy to induce, personnel and equipment safety and engineering construction progress are seriously threatened, and accurate real-time early warning of the rock explosion of the deep underground engineering is a key problem to be solved urgently in the field of the deep underground engineering safety. At present, the engineering practice mainly depends on a rock burst early warning technology system taking microseism monitoring as a core, but the existing method has remarkable limitations in timeliness, accuracy and intelligentization level, and is specifically expressed as follows: (1) The traditional method has strong subjectivity and low efficiency, and is difficult to meet the real-time early warning requirement. In the prior art, the microseismic signals are manually interpreted highly depending on expert experience, and comprehensive analysis is required by combining time domain, frequency domain and amplitude characteristics. The process is time-consuming and labor-consuming, introduces significant time delay, and has poor consistency of interpretation results due to strong subjectivity, thereby severely restricting the instantaneity and reliability of the early warning system. (2) The traditional model has limited mechanism, and is difficult to capture nonlinear dynamic characteristics of the rock burst inoculation process. The early warning method based on the experience index criterion or the shallow statistical model is simple and easy to implement, but is difficult to describe the highly nonlinear and non-stable time sequence evolution rules presented by stress accumulation, crack initiation and expansion in the rock burst inoculation process. The numerical simulation method has inherent limitations in the aspects of reproducing dynamic characteristics such as rock burst, energy release and the like due to the applicability of constitutive relations, parameter uncertainty and dependence on a small deformation theory. The method is mostly suitable for long-term risk assessment, and is difficult to deal with dynamic evolution of construction period risks. (3) The feature extraction of the existing intelligent model is incomplete, and critical frequency domain precursor information is ignored. While prior inventions have attempted to introduce deep learning models such as long and short term memory networks, convolutional neural networks, etc. to enhance predictive capabilities, most of these models focus on analysis from a temporal or spatial perspective. The microseismic signal is used as a non-stationary time sequence, and the frequency domain characteristic is a key criterion for identifying the rock burst precursor. The existing model is generally lack of an efficient frequency domain feature extraction module, so that precursor information borne by frequency components is underutilized, and further improvement of prediction accuracy is limited. (4) The model has insufficient robustness and single prediction dimension. The construction site environment of the deep underground engineering is complex, and the microseismic signals are often mixed with strong interferences such as blasting vibration, mechanical noise and the like. The existing model generally lacks a targeted noise suppression and self-adaptive characteristic enhancement mechanism, so that the generalization capability of the model is poor under the condition of low signal to noise ratio, and the false alarm rate is high. In addition, most methods can only output the overall risk level, and cannot realize the joint prediction and the fine classification of the multidimensional feature quantity of the rock burst occurrence time, the spatial position and the intensity, so that the accurate prevention and control decision of 'time-space-intensity' integration is difficult to support. In summary, the prior art lacks an intelligent early warning scheme capable of efficiently fusing time-frequency domain features, adaptively suppressing noise and accurately realizing the combined prediction of rock burst multidimensional features and risk levels, and the defect becomes a technical bottleneck for restricting the improvement of rock burst disaster prevention and control capability of deep underground engineering. Disclosure of Invention Aiming at the problems existing in the prior art, the invention provides an intelligent early warning model and method integrating frequency domain feature extraction, a channe