CN-121993257-A - Rock burst intelligent early warning method based on cross-modal fusion of while-drilling sensing data and microseismic monitoring data

Abstract

The invention provides a rock burst intelligent early warning method based on cross-modal fusion of while-drilling sensing data and microseismic monitoring data, which belongs to the technical field of tunnel and underground engineering safety and disaster prevention and control, and comprises the steps of acquiring while-drilling parameters in real time and recording three-dimensional track coordinates of drilling holes; the method comprises the steps of collecting microseismic event signals induced by drilling disturbance in real time, inverting seismic source parameters, preprocessing and aligning the collected drilling parameters and the microseismic event signals in time and space to generate an aligned multi-modal feature sequence, inputting the aligned multi-modal feature sequence into a trained rock burst risk prediction model, outputting rock burst risk level probabilities distributed along the axes of drilling tracks, carrying out spatial interpolation on the rock burst risk level probabilities distributed along the axes of the drilling tracks to generate a three-dimensional rock burst risk probability field in front of a tunnel face, carrying out three-dimensional visual reconstruction on the three-dimensional rock burst risk probability field, carrying out risk level discrimination according to the three-dimensional risk cloud picture in front of the tunnel face, and automatically issuing corresponding early warning signals.

Inventors

• DAI FENG
• LIANG BIN
• LIU YI
• WEI MINGDONG
• WU YANLIANG

Assignees

• 四川大学

Dates

Publication Date

20260508

Application Date

20260128

Claims (4)

1. 1. The rock burst intelligent early warning method based on cross-modal fusion of while-drilling sensing data and microseismic monitoring data is characterized by comprising the following steps: real-time obtaining drilling parameters such as drilling pressure, torque, propulsion speed, rotating speed and the like by using a drilling parameter sensing unit which is arranged on a drill rod, a drill bit and a drill boom of drilling and blasting construction equipment, and recording three-dimensional track coordinates of a drilling hole; Microseismic event signals induced by drilling disturbance are acquired in real time by utilizing microseismic sensors arranged in the face and surrounding rock drilling holes, and seismic source parameters such as three-dimensional space positions, seismic magnitudes, energy, frequency spectrum characteristics and the like of a seismic source are inverted; preprocessing and space-time alignment are carried out on the acquired while-drilling parameters and the microseismic event signals, and an aligned multi-mode feature sequence is generated, wherein the aligned multi-mode feature sequence comprises an aligned one-dimensional while-drilling feature sequence and an aligned one-dimensional microseismic feature sequence; Inputting the aligned multi-modal feature sequences into a trained rock burst risk prediction model, and outputting rock burst risk level probabilities distributed along the axes of all drilling tracks; Spatial interpolation is carried out on rock burst risk level probabilities distributed along the axes of all drilling tracks, a three-dimensional rock burst risk probability field in front of a tunnel face is generated, three-dimensional visual reconstruction is carried out on the three-dimensional rock burst risk probability field, risk level discrimination is carried out according to a three-dimensional risk cloud picture in front of the tunnel face, and corresponding early warning signals are automatically issued.
2. 2. The intelligent rock burst early warning method based on cross-modal fusion of while-drilling sensing data and microseismic monitoring data according to claim 1, wherein the preprocessing and space-time alignment of the acquired while-drilling parameters and microseismic event signals comprises the following steps: Filtering the while-drilling parameter sequence to obtain a smooth and reliable while-drilling parameter sequence; Noise separation and interference event rejection are carried out on the microseism event signals, and real microseism event signals are obtained; normalizing the processed while-drilling parameters and microseismic event signals to eliminate dimensional differences among different parameters; Drilling response characteristics such as the drilling pressure fluctuation rate, the torque mutation coefficient and the like are calculated based on the drilling parameter sequence, and rock fracture characteristics such as the seismic source space concentration degree, the energy index, the b value and the like are calculated based on the microseism event signal; Three-dimensional trajectory of the borehole Micro-seismic source Under the condition that the positions are unified to a global coordinate system of tunnel construction, time synchronization and checking are carried out on all data acquisition equipment, so that all acquired data are ensured to have unified and standardized time records To ensure the accuracy of the subsequent space-time correlation; By drilling three-dimensional trajectories Is a one-dimensional reference axis, and is arranged at a fixed space step distance Dividing it into T consecutive spatial segments According to the track data and the unified time Calibrating the drill bit to drill each space segment Start-stop time of (a) ; For any space segment Acquiring start-stop time All the parameters acquired in the drill bit are used for calculating the fluctuation rate of the drill bit and the torque mutation coefficient according to the parameters, and the segmented feature vector while drilling is generated ; For the same space segment The following spatio-temporal correlation and aggregation steps are performed: (1) Determining a spatial range and a temporal range; In space segments The corresponding drilling track axis is taken as the center, and a cylindrical space neighborhood with a preset radius R is determined to be a space range; To start and stop time Is a time range; (2) Screening microseismic events; Screening out microseismic events meeting the following two conditions simultaneously to form a microseismic event set : Condition one the source location of an event Located in space segments Is within the spatial range of (2); Condition two, time of occurrence of event Located in space segments Is within a time range of (2); (3) Aggregating event features; if the microseismic event is gathered If the vibration is empty, the micro-vibration characteristic vector is obtained Assigning a zero vector; if the microseismic event is gathered If not, carrying out statistical calculation on all event features in the set to obtain a microseismic feature vector ; Segmenting each space Corresponding feature vector while drilling And microseismic feature vector Combining the two feature pairs to generate two one-dimensional feature sequences while drilling which are strictly aligned on space segmentation And one-dimensional microseismic feature sequences ; If the while-drilling data at a certain position is missing, generating a substitution value by adopting adjacent interpolation, and synchronously outputting a quality mask And sample weights So as to automatically reduce the weight in the risk prediction module.
3. 3. The intelligent rock burst early warning method based on cross-modal fusion of while-drilling perception data and microseismic monitoring data according to claim 2, wherein the rock burst risk prediction model adopts a deep learning model fused with a double-branch independent feature extraction network, a cross-modal transducer fusion unit and a risk level prediction unit, and the construction and training processes are as follows: Collecting historical drilling parameters and microseismic event signals during tunnel drilling operation, preprocessing and aligning in time and space, constructing a multi-mode labeling data set by using a historical aligned multi-mode characteristic sequence and a corresponding rock burst risk level label, and dividing the multi-mode labeling data set into a training set, a verification set and a test set according to a preset proportion; the double-branch independent feature extraction network construction consists of an information while-drilling feature extraction branch and a microseismic data feature extraction branch, and the specific construction process is as follows: The while-drilling information feature extraction branch consists of a mixed architecture of a one-dimensional convolutional neural network and a transducer encoder, the one-dimensional convolutional neural network adopts a plurality of residual convolution units to perform feature extraction on local mutation and short range modes of while-drilling parameters in an input one-dimensional while-drilling feature sequence, the local feature sequence extracted by the one-dimensional convolutional neural network is input into the transducer encoder, the self-attention mechanism is utilized to establish long range relevance of local features on the whole drilling depth sequence so as to represent the overall distribution and change rule of the mechanical properties of the rock mass, and finally a group of high-dimensional while-drilling feature vector sequences F d are output; the microseismic data feature extraction branch processes the input one-dimensional microseismic feature sequence through a full connection layer to obtain a high-dimensional microseismic feature representation Calculating the association weights among the microseismic activity states at different moments in the one-dimensional microseismic feature sequence by using an independent transducer encoder and utilizing a self-attention mechanism of the transducer encoder, and finally outputting a group of weighted microseismic feature vector sequences F s ; The cross-modal transducer performs bidirectional information interaction by adopting a cross-modal attention unit, and stacks four layers of interaction units to obtain a depth-fused feature vector sequence The specific construction process is as follows: (a) Bidirectional interaction, namely micro-seismic query while drilling; In one interaction direction, in the micro-vibration characteristic sequence As Query to feature sequence while drilling Respectively used as a Key Key and a Value, and comprises the following calculation processes: ; Wherein, the In order for the projection matrix to be learnable, Reliability marking by data processing modules Constructing; (b) Bidirectional interaction, namely inquiring microseismic while drilling; In another interaction direction, with feature sequence while drilling As Query, the microseismic characteristic sequence Respectively used as a Key Key and a Value, and comprises the following calculation processes: ; Wherein, the Is a learnable projection matrix; (c) Form of attention calculation; For any attention head, let the attention score matrix be: ; Wherein, the In the single-head dimension, the device has the advantages of high precision, In order to Query the Query matrix, Is a Key Key matrix; Cross-modal attention The following formula is adopted: ; Wherein, the Is a one-dimensional position mask vector, Is a preset positive number, and V is a Value matrix; (d) Feature fusion and transformation; Will be And (3) with And then a feedforward neural network FFN is fed, and residual error connection and layer normalization are combined, so that dimensional difference caused by feature splicing is solved, and a linear projection layer is introduced to perform dimensional alignment on a residual error path: ; ; Wherein, the The channel splice is indicated as such, A linear projection matrix of the residual path for the spliced original input features Mapping to and The dimensions of the same are such that, Is a fully connected network of two layers, Is a weight matrix of the linear projection layer, Is the weight bias vector of the linear projection layer, Pressing the spliced dimension back to the target dimension The training stability is improved by matching residual error connection with layer normalization; By stacking four layers of cross-attention computing units, information of two modes is repeatedly interacted and refined on different layers, and finally a depth-fused feature vector sequence is output ; The risk level prediction unit consists of a global average pooling layer and at least one full connection layer, and the specific construction process is as follows: the global average pooling layer pools feature vector sequences Average aggregation along the length dimension of the sequence to generate a single feature vector with fixed dimension ; Will be Firstly, a full-connection layer with an activation function is sent into, and then, a full-connection output layer is sent into, so that un-normalized scores of all risk levels are obtained: ; Wherein, the As the number of risk classes to be used, For the nonlinear activation function ReLU, And For the weight matrix and bias vector of the full connection layer with activation function, And The weight matrix and the bias vector are the weight matrix and the bias vector of the full-connection output layer; the score of the last step is outputted through the Softmax probability Conversion to a hierarchical probability distribution vector : ; Wherein, the Indicating occurrence of rock mass Predictive probability of a stage rock burst, and ; Model training is carried out by adopting a weighted cross entropy loss function, an Adam optimizer and a Dropout regularization technology, and the specific steps are as follows: Taking the while-drilling feature sequences and the microseismic feature sequences in the training set, the verification set and the test set as model input, and taking the rock burst risk level labels as supervision signals to form training pairs for supervision learning; employing weighted cross entropy as the primary penalty and introducing sample weights from the data processing module The method specifically comprises the following steps: ; Wherein, the In the case of a batch size of the product, For the length of the sequence, As a total number of risk levels, For the class weight to be a class weight, For the sample In position Is used for the weight of the (c), As a real tag it is possible to provide a real tag, For the model prediction output, Is a one-hot tag which is used for the identification of a user, ; Adopting an Adam optimizer to iteratively update all the trainable parameters, calculating gradients through back propagation, and carrying out weight updating by combining first-order/second-order momentum estimation until a loss function converges or reaches a preset training round upper limit; Dropout regularization is introduced after the FFN layers of the full connection layer, the transducer encoder and the fusion module, and the training is performed with preset probability The partial neuron outputs are randomly zeroed, so that the model learns more robust characterization.
4. 4. The intelligent rock burst early warning method based on cross-modal fusion of perception data while drilling and microseismic monitoring data according to claim 1, wherein the rock burst risk level probability distributed along each drilling track axis is spatially interpolated to generate a three-dimensional rock burst risk probability field in front of a tunnel face, the three-dimensional rock burst risk probability field is subjected to three-dimensional visual reconstruction, risk level discrimination is performed according to the three-dimensional rock burst risk probability field in front of the tunnel face, and corresponding early warning signals are automatically issued, and the intelligent rock burst early warning method comprises the following steps: expanding the rock burst risk level probability distributed along the axis of each drilling track to a three-dimensional space of an un-drilled area in front of the whole face by a kriging interpolation method, gridding a target space with preset precision, calculating the risk level probability of each grid node, and finally constructing a continuous three-dimensional rock burst risk probability field formed by grid units; Constructing a three-dimensional basic model containing tunnel design outline, excavated space and geological structure information, carrying out space registration and fusion on a continuous three-dimensional rock burst risk probability field and the three-dimensional basic model, and endowing each grid unit in the continuous three-dimensional rock burst risk probability field with specific color and transparency according to the risk level probability by setting a color and transparency mapping function to form a dynamically updated three-dimensional risk cloud picture on the three-dimensional basic model; The method comprises the steps of presetting four-level risk probability thresholds corresponding to rock burst risk levels, namely a blue early warning threshold, a yellow early warning threshold, an orange early warning threshold and a red early warning threshold, searching the maximum risk probability value in a continuous three-dimensional rock burst risk probability field in real time at fixed frequency, and triggering early warning signals of corresponding levels when the maximum risk probability value of any grid unit meets a preset threshold condition of a certain level.

Description

Rock burst intelligent early warning method based on cross-modal fusion of while-drilling sensing data and microseismic monitoring data Technical Field The invention relates to the technical field of tunnel and underground engineering safety and disaster prevention, in particular to a rock burst intelligent early warning method based on cross-modal fusion of while-drilling sensing data and microseismic monitoring data. Background Along with the continuous promotion of deep tunnel and underground engineering construction, the construction section is generally in complex environments such as high ground stress, high osmotic pressure and strong disturbance, and rock burst disasters are frequent, and the construction section becomes one of main safety risks of deep-buried hard rock tunnel construction. The drilling and blasting method is a main construction method for deep tunnel excavation, and the strong power disturbance generated in the operation process is extremely easy to induce rock blasting, so that serious threat is caused to the safety of constructors and the operation of equipment. Therefore, the accurate advanced quantitative pre-judgment of the potential rock burst risk of surrounding rock in front of the face is carried out before the blasting operation, and is a key technical problem to be solved in the prior art for realizing the safety prevention and control of rock burst disasters. At present, a microseismic monitoring technology is widely applied to identification and early warning of rock burst risk, and the technology collects elastic wave signals generated by rock mass fracture through sensors arranged in surrounding rock and inverts seismic source parameters according to the elastic wave signals to evaluate the risk. Although the microseismic monitoring technology is mature in the aspects of signal acquisition and analysis, the microseismic monitoring technology is passive in nature, can only send out an alarm after the inside of a rock mass is damaged and energy is released, has limited identification capability and insufficient early warning timeliness for the inoculation stage of rock burst, namely the stage of high stress concentration but not macroscopic fracture, and is difficult to meet the requirement of engineering practice on advanced prejudgment. In order to make up for the defect of timeliness of single microseismic monitoring, the prior art tries to represent the surrounding rock state in front of the face by using parameters while drilling, but related applications are limited to static classification of the surrounding rock class or stability, active while drilling sensing data representing the physical properties of the rock mass are not yet deeply coupled with passive microseismic monitoring signals reflecting the damage evolution process, and a unified dynamic rock burst risk criterion is constructed. In addition, the artificial intelligence method is also gradually applied to rock burst risk prediction, but the current deep learning method mostly adopts a fusion strategy of feature splicing, and ignores the essential difference of different modal data on space-time distribution features (namely continuous linear distribution of while-drilling sensing data and discrete punctiform distribution of microseismic monitoring signals), so that a model cannot effectively establish physical mapping of the two, the fusion effect is poor, a prediction result is often disjointed with the actual operation rhythm, and on-site construction decision is difficult to directly serve. Disclosure of Invention Aiming at the problems existing in the prior art, the invention provides an intelligent early warning scheme capable of realizing advanced pre-judging of rock burst risk, wherein while-drilling perception is introduced as an active detection means, and the intelligent early warning scheme forms space-time complementation with a passive response means of microseismic monitoring, discrete microseismic events and continuous while-drilling parameters are unified into the same space-time reference system, two types of time-space synchronous characteristic sequence pairs are generated, intrinsic correlation characteristics of the discrete microseismic events and the continuous while-drilling parameters in the rock burst inoculation process are revealed by using a cross-modal deep learning model, and finally a prediction result is visualized into a three-dimensional rock burst risk cloud picture capable of directly guiding construction decisions. The invention provides a rock burst intelligent early warning method based on cross-modal fusion of while-drilling sensing data and microseismic monitoring data, which comprises the following steps: real-time obtaining drilling parameters such as drilling pressure, torque, propulsion speed, rotating speed and the like by using a drilling parameter sensing unit which is arranged on a drill rod, a drill bit and a drill boom of drilling and blasting con