A system and method for real-time monitoring and early warning of surrounding rock stress while drilling
By integrating multi-source information processing of drilling parameters, drill cuttings information, and microseismic signals, and combining it with an LSTM deep recurrent neural network, the problem of insensitivity in traditional surrounding rock stress monitoring has been solved. This enables dynamic, real-time, and multi-parameter comprehensive analysis of coal mine drilling operations, improving safety and intelligence levels.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF MINING & TECH
- Filing Date
- 2026-01-14
- Publication Date
- 2026-05-26
Smart Images
Figure CN122090584A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of coal mine safety monitoring technology, specifically to a real-time monitoring and early warning system and method for surrounding rock stress during drilling. Background Technology
[0002] In coal mine roadways, the stress state of the surrounding rock directly affects the safety of drilling operations. Traditional methods of monitoring surrounding rock stress often rely on monitoring drilling parameters to infer formation information. This approach, dependent solely on drilling parameters, is susceptible to interference from coal and rock drilling conditions. It also suffers from problems such as insensitivity to sudden changes in surrounding rock stress, response lag, and a single monitoring dimension, making it difficult to achieve dynamic, real-time, and multi-parameter comprehensive analysis during the drilling process. Therefore, a method is needed that can integrate real-time acquisition, processing, and fusion analysis of multi-source information from drilling parameters, drill cuttings information, and microseismic signals. Furthermore, an intelligent early warning model with time-series modeling capabilities is required to achieve dynamic, graded early warning of the surrounding rock stress state during drilling, thereby improving the safety and intelligence level of coal mine drilling operations. Summary of the Invention
[0003] In view of the above-mentioned technical deficiencies, the purpose of this invention is to provide a real-time monitoring and early warning system and method for surrounding rock stress during drilling, so as to improve the safety and intelligence level of coal mine drilling operations.
[0004] To achieve the above objectives, the present invention adopts the following technical solution:
[0005] A real-time monitoring and early warning system for surrounding rock stress during drilling includes a drilling information acquisition module, a drilling information processing module, a video processing module, a microseismic information processing module, and a comprehensive information processing and early warning module.
[0006] The drilling information acquisition module includes pressure sensor one, pressure sensor two, displacement sensor, intrinsically safe camera for mining, and micro-vibration sensor; pressure sensor one and pressure sensor two are respectively installed on the feed cylinder and the circulating hydraulic pipeline of the hydraulic rotary motor of the drilling rig; displacement sensor is installed on the piston inside the hydraulic cylinder; intrinsically safe camera for mining is installed on one side of the borehole at the same level; micro-vibration sensor is installed in the coal seam on one side away from the borehole;
[0007] The drilling information processing module is used to receive the feed pressure signal, rotation pressure signal and feed cylinder extension and retraction displacement signal output by pressure sensor 1, pressure sensor 2 and displacement sensor, and to perform noise reduction processing.
[0008] The video processing module is used to receive drill cuttings image data at the borehole location output by the intrinsically safe mining camera, and to perform noise reduction and image feature extraction on the data;
[0009] The microseismic information processing module is connected to the microseismic sensor and is used to receive microseismic sensor data, perform wavelet transform denoising processing, and calculate microseismic energy.
[0010] The integrated information processing and early warning module is connected to the drilling information processing module, video processing module and microseismic information processing module. It fuses the input multi-source data and inputs the fused data into the surrounding rock stress monitoring model, and outputs monitoring and early warning results.
[0011] A method for real-time monitoring and early warning of surrounding rock stress during drilling includes the following steps:
[0012] S1: Real-time acquisition of drilling parameters, drill cuttings image data, and micro-vibration signals via the information acquisition module;
[0013] S2: The collected data are processed by the drilling information processing module, video processing module, and microseismic information processing module for denoising and feature extraction.
[0014] S3: The multi-source data is fused through the integrated information processing and early warning module to obtain a fused dataset;
[0015] S4: Input the fused dataset into the surrounding rock stress monitoring model and output the early warning result of the surrounding rock stress hazard level;
[0016] S5: Regularly update the core parameters of the surrounding rock stress monitoring model to improve the accuracy of early warning.
[0017] Preferably, in step S2, the processing procedure of the drilling parameters by the drilling information processing module is as follows:
[0018] The drilling parameters are processed by sliding translation filtering, and the output feed pressure signal is: Rotary pressure signal Input hydraulic cylinder extension / retraction displacement signal , It is a sequence number related to time, the filter window is defined as L, and the output processed feed pressure signal is calculated according to formula (1). : ;
[0019] in, This is a summation index variable used to iterate through all integers from 0 to L-1;
[0020] The processed rotary pressure signal is calculated using formula (1). and the extension / retraction displacement signal of the feed cylinder The denoised drilling parameters are represented in the dataset as follows: .
[0021] Preferably, in step S2, the video processing module processes the drill cuttings data as follows:
[0022] The video output from the camera is frame-by-frame extracted at 10ms intervals; the drill cuttings flow rate and particle size images of each extracted frame are processed using Gaussian blur denoising; the Gaussian convolution kernel is... The standard deviation is The origin (0,0) is at the center of the kernel. Substituting the image coordinates (x,y) into the two-dimensional Gaussian function formula (2), we calculate the weight value at each position in the kernel: ;
[0023] Normalize the original weight values; slide this Gaussian kernel across the original image to perform convolution calculation; output each pixel in the image for each convolution calculation using formula (3): ;
[0024] in, It is the pixel value at the corresponding position in the original frame-sampling image; This is the weight value of the Gaussian convolution kernel at position (i,j); the convolution calculation needs to be performed independently on each channel of the frame-sampled image;
[0025] The denoised frame-by-frame images are input into a CNN feature extraction model. Multiple sequentially stacked convolutional blocks are used for convolution calculations, non-linear activation, and spatial downsampling to obtain feature vectors representing the drill cuttings flow rate and granularity of the frame-by-frame images. The drill cuttings dataset after frame extraction, denoising, and feature vector extraction is represented as follows: .
[0026] Preferably, in step S2, the microseismic information processing module processes the microseismic information as follows:
[0027] 1) Original acquired microseismic waveform signal The discrete signal is obtained by sampling using formula (4). sampling frequency ,in , Signal length: ;
[0028] in, Is t= The unit impulse function at time t;
[0029] 2) Determine the wavelet decomposition coefficients;
[0030] Decomposition layer number For each layer The approximation coefficients of the j-th layer are calculated using formula (5). The detail factor of the j-th layer is calculated using formula (6). : ; ;
[0031] in, It is a low-pass filter that preserves the low-frequency components of microseismic signals. It is a high-pass filter that preserves the high-frequency components of the microseismic signal. n-2k is the relative displacement of the filter, ensuring that only a continuous time interval of the associated signal is calculated each time. It is the discrete microseismic signal of the (j-1)th layer;
[0032] For each layer of detail coefficients Thresholding is performed using the hard thresholding function of formula (7) to obtain the processed detail coefficients. : ;
[0033] in, It is the first The threshold of the layer, , It is the first Noise standard deviation of layer detail factor;
[0034] 3) Calculate the reconstructed signal after wavelet transform;
[0035] From recursive formula (8) to This refers to the microseismic signal obtained using wavelet transform denoising. ; ;
[0036] in, For the corresponding Micro-vibration low-frequency reconstruction filter; For the corresponding Micro-vibration high-frequency reconstruction filter;
[0037] 4) Calculate the microseismic energy from the denoised microseismic signal using formula (9): ;
[0038] in, R is the density of coal and rock at the installation location of the microseismic sensor; R is the distance from the borehole to the microseismic sensor. These are the particle velocities of the P-wave and S-wave bands separated from the waveforms recorded by the microseismic sensor, respectively. The P-wave and S-wave velocities of coal and rock at the installation location of the microseismic sensor; , The propagation time of the P wave and S wave;
[0039] No. Maximum microseismic energy within a time window: ;
[0040] No. Average microseismic energy within a time window: .
[0041] Preferably, in step S3, the integrated information processing and early warning module performs the following fusion processing of multi-source data:
[0042] Merging datasets It is calculated using a weighted average algorithm of drilling parameters and drill cuttings data;
[0043] By integrating drilling parameters, drill cuttings data, and microseismic data according to the same timestamp, the average value of the data in the i-th time window is calculated, denoted as: The variances of the corresponding data are respectively , ,…, ;
[0044] Using formula (10), the weight values corresponding to the minimum total mean square error are obtained: ;
[0045] This corresponds to the minimum total mean square error. We can calculate the result using formula (11): ;
[0046] Merging datasets ;in These are the weights of each data point, and their sum is 1.
[0047] Merge datasets Input the surrounding rock stress monitoring and early warning model, and output the monitoring results.
[0048] Preferably, in step S4, the surrounding rock stress monitoring model is based on an LSTM deep recurrent neural network, and the construction and early warning process is as follows:
[0049] The input to the monitoring model is the fused dataset extracted within each time window. Construct a time series dataset, with each window represented as a vector. ;in, The time windows are numbered; the entire dataset is represented as ;
[0050] The historical data is used to create a training label set, which is represented as follows: The time series is then divided into overlapping or non-overlapping subsequences, represented as follows: ,in This is the sampling step size; Given the length of the precursor sequence, the precursor pattern set can be obtained, represented as... ;
[0051] Will The data is fed into an LSTM for implicit feature extraction, and then average pooling is used to obtain the implicit feature vector. ;in For the corresponding The implicit features of each vector; the training label set is used as input, and its output is divided into three levels, corresponding to strong stress hazard, medium stress hazard, and weak stress hazard respectively;
[0052] The fixed-dimensional implicit features extracted by LSTM are mapped to a dimension with the same number of classes using formula (12): ;
[0053] in, The implicit feature vector extracted by LSTM; This is the weight matrix; It is the bias vector; Scores for unnormalized categories;
[0054] Classification is performed using the Softmax activation function (13): ;
[0055] in, ; For stress monitoring category, These correspond to high stress hazard, medium stress hazard, and low stress hazard, respectively.
[0056] Finally, normalization is performed to obtain the output result:
[0057] 1) Indicators: Pressure signal increases sharply or fluctuates violently, drill cuttings flow rate far exceeds the normal amount and the particle size is fine, and the maximum energy of micro-vibration is >10. 5 J; Hazard level: High stress hazard;
[0058] 2) Indicators: Increased or drastic pressure signal, increased drill cuttings flow rate and smaller particle size, and maximum micro-vibration energy of 10. 2 -10 5 J; Hazard level: Moderate stress hazard;
[0059] 3) Indicators: Pressure signal fluctuates slightly; drill cuttings flow rate and particle size variation range is within 5% above and below normal drilling; maximum micro-vibration energy is 10. 2 -10 4 J; Hazard level: Weak stress hazard.
[0060] Preferably, in step S5, the update process of the core parameters of the surrounding rock stress monitoring model is as follows:
[0061] In the loss calculation, the true class label of each sample is y, which is converted into a one-hot encoding Y. The loss function is expressed as: ;
[0062] in, Is the i-th sample in the true category? The predicted probability; It is the average loss of N samples in the same batch;
[0063] The gradient of the loss function with respect to the unnormalized class score z of the Softmax input: ;
[0064] in, Is the i-th sample in the true category? The predicted probability; N is the number of samples in the same batch;
[0065] In backpropagation, the loss is applied to the weights and bias vectors as well as the LSTM hidden vectors. The gradient is expressed as: ; ;
[0066] in It is the first The unnormalized score vector of each sample output; ;
[0067] Then use gradient descent to update the parameters;
[0068] Weight update: ;
[0069] in, The updated weights;
[0070] Bias vector update: ;
[0071] in, The learning rate in gradient descent. This is the updated bias vector.
[0072] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0073] This invention utilizes methods such as sliding window filtering, Gaussian blur denoising, CNN feature extraction, and wavelet denoising to process and integrate three types of heterogeneous data—drilling parameters, drill cuttings information, and microseismic signals—in real time. It constructs a multi-source information fusion analysis system and combines it with an LSTM deep recurrent neural network to model the multi-source time-series data, effectively capturing precursory features in the process of surrounding rock stress changes and achieving dynamic intelligent early warning of surrounding rock stress. By periodically updating the core parameters of the monitoring model, the accuracy of subsequent surrounding rock stress prediction is improved. Attached Figure Description
[0074] Figure 1 This is a flowchart illustrating the real-time monitoring and early warning system for surrounding rock stress during drilling, as presented in this invention. Detailed Implementation
[0075] The invention will now be further described with reference to the accompanying drawings.
[0076] like Figure 1 As shown, a real-time monitoring and early warning system for surrounding rock stress during drilling includes an information acquisition module, a drilling information processing module, a video processing module, a microseismic information processing module, and a comprehensive information processing and early warning module.
[0077] The drilling information acquisition module includes pressure sensor one, pressure sensor two, displacement sensor, intrinsically safe mining camera, and micro-vibration sensor. Pressure sensor one and pressure sensor two are respectively installed on the circulating hydraulic lines of the feed cylinder and hydraulic rotary motor of the drilling rig; the displacement sensor is installed on the piston inside the hydraulic cylinder; the intrinsically safe mining camera is installed on the left side of the borehole at the same horizontal plane, at a 45° angle to the coal wall; the micro-vibration sensor is installed in the coal seam 1.5 meters away from the borehole.
[0078] The drilling information processing module is used to receive the feed pressure signal, rotation pressure signal and feed cylinder extension and retraction displacement signal output by pressure sensor 1, pressure sensor 2 and displacement sensor, and to perform noise reduction processing.
[0079] The video processing module is used to receive drill cuttings image data at the borehole location output by the intrinsically safe mining camera, and to perform noise reduction and image feature extraction on the data;
[0080] The microseismic information processing module is connected to the microseismic sensor and is used to receive microseismic sensor data, perform wavelet transform denoising processing, and calculate microseismic energy.
[0081] The integrated information processing and early warning module is connected to the drilling information processing module, video processing module and microseismic information processing module. It fuses the input multi-source data and inputs the fused data into the surrounding rock stress monitoring model, and outputs the monitoring and early warning results.
[0082] As a further optimization scheme, the processing procedures for the feed pressure signal, rotary pressure signal, and feed cylinder extension / retraction displacement signal in the drilling information processing module are as follows:
[0083] After performing sliding translation filtering, the output feed pressure signal is: Rotary pressure signal Input hydraulic cylinder extension / retraction displacement signal , It is a sequence number related to time, the filter window is defined as L, and the output processed feed pressure signal is calculated according to formula (1). : ;
[0084] in, This is a summation index variable used to iterate through all integers from 0 to L-1;
[0085] Similarly, the slewing pressure signal is calculated using formula (1) and then output. and the extension / retraction displacement signal of the feed cylinder The denoised drilling parameters are represented in the dataset as follows: ;
[0086] As a further optimization, the processing procedure for drill cuttings data in the video processing module is as follows:
[0087] The video output by the camera is subjected to frame extraction at equal intervals of 10ms / time;
[0088] The cuttings flow rate and particle size images of each frame were processed using Gaussian blur denoising.
[0089] Gaussian convolution kernel is The standard deviation is The origin (0,0) is at the center of the kernel. Substituting the image coordinates (x,y) into the two-dimensional Gaussian function formula (2), we calculate the weight value at each position in the kernel: ;
[0090] Normalize the original weight values; slide this Gaussian kernel across the original image to perform convolution calculation. The convolution calculation outputs each pixel in the image for each iteration using formula (3): ;
[0091] in, It is the pixel value at the corresponding position in the original frame-sampling image; This is the weight value of the Gaussian convolution kernel at position (i,j); the convolution calculation needs to be performed independently on each channel of the frame-sampled image;
[0092] As a further optimization, the denoised frame-by-frame images are input into a CNN feature extraction model. Multiple sequentially stacked convolutional blocks are used for convolution calculations, non-linear activation, and spatial downsampling to obtain feature vectors representing the drill cuttings flow rate and granularity of the frame-by-frame images. ;
[0093] The drill cuttings dataset after frame extraction, denoising, and feature vector extraction is represented as follows: ;
[0094] As a further optimization, the microseismic information processing process in the microseismic information processing module is as follows:
[0095] 1) Original acquired microseismic waveform signal The discrete signal is obtained by sampling using formula (4). sampling frequency ,in , Signal length: ;
[0096] in, Is t= The unit impulse function at time t;
[0097] 2) Determine the wavelet decomposition coefficients;
[0098] Decomposition layer number For each layer The approximation coefficients of the j-th layer are calculated using formula (5). The detail factor of the j-th layer is calculated using formula (6). : ; ;
[0099] in, It is a low-pass filter that preserves the low-frequency components of microseismic signals. It is a high-pass filter that preserves the high-frequency components of the microseismic signal. n-2k is the relative displacement of the filter, ensuring that only a continuous time interval of the associated signal is calculated each time. It is the discrete microseismic signal of the (j-1)th layer;
[0100] For each layer of detail coefficients Thresholding is performed using the hard thresholding function of formula (7) to obtain the processed detail coefficients. : ;
[0101] in, It is the first The threshold of the layer, , It is the first Noise standard deviation of layer detail factor;
[0102] 3) Calculate the reconstructed signal after wavelet transform;
[0103] From recursive formula (8) to This refers to the microseismic signal obtained using wavelet transform denoising. ; ;
[0104] in, For the corresponding Micro-vibration low-frequency reconstruction filter; For the corresponding Micro-vibration high-frequency reconstruction filter;
[0105] 4) Calculate the microseismic energy from the denoised microseismic signal using formula (9): ;
[0106] in, R is the density of coal and rock at the installation location of the microseismic sensor; R is the distance from the borehole to the microseismic sensor. These are the particle velocities of the P-wave and S-wave bands separated from the waveforms recorded by the microseismic sensor, respectively. The P-wave and S-wave velocities of coal and rock at the installation location of the microseismic sensor; , The propagation time of the P wave and S wave;
[0107] No. Maximum microseismic energy within a time window: ;
[0108] No. Average microseismic energy within a time window: .
[0109] The fusion processing of drilling parameters, drill cuttings data, and microseismic data in the integrated information processing and early warning module is as follows:
[0110] Merging datasets It is calculated using a weighted average algorithm of drilling parameters and drill cuttings data;
[0111] By integrating drilling parameters, drill cuttings data, and microseismic data according to the same timestamp, the average value of the data in the i-th time window is calculated, denoted as: The variances of the corresponding data are respectively , ,…, ;
[0112] Using formula (10), the weight values corresponding to the minimum total mean square error are obtained: ;
[0113] This corresponds to the minimum total mean square error. We can calculate the result using formula (11): ;Merge datasets ;in These are the weights of each data point, and their sum is 1.
[0114] Merge datasets Input the surrounding rock stress monitoring and early warning model, and output the monitoring results;
[0115] The early warning mechanism in the integrated information processing and early warning module relies on constructing an LSTM deep recurrent neural network model for monitoring surrounding rock stress. The model construction method is as follows:
[0116] The input to the monitoring model is the fused dataset extracted within each time window. Construct a time series dataset, with each window represented as a vector. ,in, The time windows are numbered; the entire dataset can be represented as ;
[0117] The historical data is used to create a training label set, which is represented as follows: The time series is then divided into overlapping or non-overlapping subsequences, represented as follows: ,in This is the sampling step size; Given the length of the precursor sequence, the precursor pattern set can be obtained, represented as... ;
[0118] Will The data is fed into an LSTM for implicit feature extraction, and then average pooling is used to obtain the implicit feature vector. ,in For the corresponding The implicit features of each vector; the training label set is used as input, and its output is divided into three levels, corresponding to strong stress hazard, medium stress hazard, and weak stress hazard respectively;
[0119] The fixed-dimensional implicit features extracted by LSTM are mapped to a dimension with the same number of classes using formula (12): ;
[0120] in, The implicit feature vector extracted by LSTM; This is the weight matrix; It is the bias vector; Scores for unnormalized categories;
[0121] Classification is performed using the Softmax activation function (13): ;
[0122] in, , For stress monitoring category, These correspond to high stress hazard, medium stress hazard, and low stress hazard, respectively.
[0123] Finally, normalization is performed to obtain the output result:
[0124] index Hazard level <![CDATA[The pressure signal increases sharply or fluctuates violently, the drill cuttings flow rate far exceeds the normal value and the particle size is fine, and the maximum energy of microseismicity > 10 5 J]]> High stress hazard <![CDATA[The pressure signal increases or fluctuates violently, the drilling cuttings flow rate increases and the particle size becomes smaller, and the maximum energy of microseismicity is 10 2 -10 5 J]]> Medium stress hazard <![CDATA[The pressure signal fluctuates slightly, the change range of the cuttings flow rate and particle size is within 5% above and below the normal drilling, and the maximum energy of microseismicity is 10 2 -10 4 J]]> Weak stress danger
[0125] As a further optimization approach, the core parameters are updated periodically based on the constructed surrounding rock stress monitoring model.
[0126] In the loss calculation, the true class label of each sample is y, which is converted into a one-hot encoding Y. The loss function is expressed as: ;
[0127] in, Is the i-th sample in the true category? The predicted probability; It is the average loss of N samples in the same batch;
[0128] The gradient of the loss function with respect to the unnormalized class score z of the Softmax input: ;
[0129] in, Is the i-th sample in the true category? The predicted probability; N is the number of samples in the same batch;
[0130] In backpropagation, the loss is applied to the weights and bias vectors as well as the LSTM hidden vectors. The gradient is expressed as: ; ;
[0131] in It is the first The unnormalized score vector of each sample output: ;
[0132] Then use gradient descent to update the parameters;
[0133] Weight update: ;
[0134] in, The updated weights;
[0135] Bias vector update: ;
[0136] in, The learning rate in gradient descent. This is the updated bias vector.
[0137] This invention, by processing and fusing drilling parameters, drill cuttings information, and microseismic signal data in real time, and combining this with an LSTM deep recurrent neural network to establish a monitoring model for multi-source time-series data, has significant implications for real-time monitoring of surrounding rock stress. Compared to existing technologies, this invention analyzes a wider variety of data types related to surrounding rock stress, providing a more comprehensive reflection of the stress evolution process.
Claims
1. A real-time monitoring and early warning system for surrounding rock stress during drilling, characterized in that, It includes a drilling information acquisition module, a drilling information processing module, a video processing module, a microseismic information processing module, and a comprehensive information processing and early warning module; The drilling information acquisition module includes pressure sensor one, pressure sensor two, displacement sensor, intrinsically safe camera for mining, and micro-vibration sensor; pressure sensor one and pressure sensor two are respectively installed on the feed cylinder and the circulating hydraulic pipeline of the hydraulic rotary motor of the drilling rig; displacement sensor is installed on the piston inside the hydraulic cylinder; intrinsically safe camera for mining is installed on one side of the borehole at the same level; micro-vibration sensor is installed in the coal seam on one side away from the borehole; The drilling information processing module is used to receive the feed pressure signal, rotation pressure signal and feed cylinder extension and retraction displacement signal output by pressure sensor 1, pressure sensor 2 and displacement sensor, and to perform noise reduction processing. The video processing module is used to receive drill cuttings image data at the borehole location output by the intrinsically safe mining camera, and to perform noise reduction and image feature extraction on the data; The microseismic information processing module is connected to the microseismic sensor and is used to receive microseismic sensor data, perform wavelet transform denoising processing, and calculate microseismic energy. The integrated information processing and early warning module is connected to the drilling information processing module, video processing module and microseismic information processing module. It fuses the input multi-source data and inputs the fused data into the surrounding rock stress monitoring model, and outputs monitoring and early warning results.
2. A method for real-time monitoring and early warning of surrounding rock stress during drilling, characterized in that, Includes the following steps: S1: Real-time acquisition of drilling parameters, drill cuttings image data, and micro-vibration signals via the information acquisition module; S2: The collected data are processed by the drilling information processing module, video processing module, and microseismic information processing module for denoising and feature extraction. S3: The multi-source data is fused through the integrated information processing and early warning module to obtain a fused dataset; S4: Input the fused dataset into the surrounding rock stress monitoring model and output the early warning result of the surrounding rock stress hazard level; S5: Regularly update the core parameters of the surrounding rock stress monitoring model to improve the accuracy of early warning.
3. The method for real-time monitoring and early warning of surrounding rock stress during drilling as described in claim 2, characterized in that, In step S2, the processing procedure for the drilling parameters by the drilling information processing module is as follows: The drilling parameters are processed by sliding translation filtering, and the output feed pressure signal is: Rotary pressure signal Input hydraulic cylinder extension / retraction displacement signal , It is a sequence number related to time, the filter window is defined as L, and the output processed feed pressure signal is calculated according to formula (1). : ； in, This is a summation index variable used to iterate through all integers from 0 to L-1; The processed rotary pressure signal is calculated using formula (1). and the extension / retraction displacement signal of the feed cylinder The denoised drilling parameters are represented in the dataset as follows: .
4. The method for real-time monitoring and early warning of surrounding rock stress during drilling as described in claim 2, characterized in that, In step S2, the video processing module processes the drill cuttings data as follows: The video output from the camera is frame-by-frame extracted at 10ms intervals; the drill cuttings flow rate and particle size images of each extracted frame are processed using Gaussian blur denoising; the Gaussian convolution kernel is... The standard deviation is The origin (0,0) is at the center of the kernel. Substituting the image coordinates (x,y) into the two-dimensional Gaussian function formula (2), we calculate the weight value at each position in the kernel: ； Normalize the original weight values; slide this Gaussian kernel across the original image to perform convolution calculation; output each pixel in the image for each convolution calculation using formula (3): ； in, It is the pixel value at the corresponding position in the original frame-sampling image; This is the weight value of the Gaussian convolution kernel at position (i,j); the convolution calculation needs to be performed independently on each channel of the frame-sampled image; The denoised frame-by-frame images are input into a CNN feature extraction model. Multiple sequentially stacked convolutional blocks are used for convolution calculations, non-linear activation, and spatial downsampling to obtain feature vectors representing the cuttings flow rate and granularity of the frame-by-frame images. The drill cuttings dataset after frame extraction, denoising, and feature vector extraction is represented as follows: .
5. The method for real-time monitoring and early warning of surrounding rock stress during drilling as described in claim 2, characterized in that, In step S2, the microseismic information processing module processes the microseismic information as follows: 1) Original acquired microseismic waveform signal The discrete signal is obtained by sampling using formula (4). sampling frequency ,in , Signal length: ； in, Is t= The unit impulse function at time t; 2) Determine the wavelet decomposition coefficients; Decomposition layer number For each layer The approximation coefficients of the j-th layer are calculated using formula (5). The detail factor of the j-th layer is calculated using formula (6). : ； ； in, It is a low-pass filter that preserves the low-frequency components of microseismic signals. It is a high-pass filter that preserves the high-frequency components of the microseismic signal, where n-2k is the relative displacement of the filter, ensuring that only a continuous time interval of the associated signal is calculated each time. It is the discrete microseismic signal of the (j-1)th layer; For each layer of detail coefficients Thresholding is performed using the hard thresholding function of formula (7) to obtain the processed detail coefficients. : ； in, It is the first The threshold of the layer, , It is the first Noise standard deviation of layer detail factor; 3) Calculate the reconstructed signal after wavelet transform; From recursive formula (8) to This refers to the microseismic signal obtained using wavelet transform denoising. ; ； in, For the corresponding Micro-vibration low-frequency reconstruction filter; For the corresponding Micro-vibration high-frequency reconstruction filter; 4) Calculate the microseismic energy from the denoised microseismic signal using formula (9): ； in, R is the density of coal and rock at the installation location of the microseismic sensor; R is the distance from the borehole to the microseismic sensor. These are the particle velocities of the P-wave and S-wave bands separated from the waveforms recorded by the microseismic sensor, respectively. The P-wave and S-wave velocities of coal and rock at the installation location of the microseismic sensor; , The propagation time of the P wave and S wave; No. Maximum microseismic energy within a time window: ; No. Average microseismic energy within a time window: .
6. The method for real-time monitoring and early warning of surrounding rock stress during drilling as described in claim 2, characterized in that, In step S3, the integrated information processing and early warning module performs the following fusion processing of multi-source data: Merging datasets It is calculated using a weighted average algorithm of drilling parameters and drill cuttings data; By integrating drilling parameters, drill cuttings data, and microseismic data according to the same timestamp, the average value of the data in the i-th time window is calculated, denoted as: The variances of the corresponding data are respectively , ,…, ; Using formula (10), the weight values corresponding to the minimum total mean square error are obtained: ； This corresponds to the minimum total mean square error. We can calculate the result using formula (11): ； Merging datasets ;in These are the weights of each data point, and their sum is 1. Merge datasets Input the surrounding rock stress monitoring and early warning model, and output the monitoring results.
7. The method for real-time monitoring and early warning of surrounding rock stress during drilling as described in claim 2, characterized in that, In step S4, the surrounding rock stress monitoring model is based on an LSTM deep recurrent neural network, and the construction and early warning process is as follows: The input to the monitoring model is the fused dataset extracted within each time window. Construct a time series dataset, with each window represented as a vector. ;in, The time windows are numbered; the entire dataset is represented as ; The historical data is used to create a training label set, which is represented as follows: The time series is then divided into overlapping or non-overlapping subsequences, represented as follows: ,in This is the sampling step size; Given the length of the precursor sequence, the precursor pattern set can be obtained, represented as... ; Will The data is fed into an LSTM for implicit feature extraction, and then average pooling is used to obtain the implicit feature vector. ;in For the corresponding The implicit features of each vector; taking the training label set as input, its output is divided into three levels, corresponding to strong stress hazard, medium stress hazard, and weak stress hazard respectively; The fixed-dimensional implicit features extracted by LSTM are mapped to a dimension with the same number of classes using formula (12): ； in, The implicit feature vector extracted by LSTM; This is the weight matrix; It is the bias vector; Scores for unnormalized categories; Classification was performed using the Softmax activation function (13); ； in, ; For stress monitoring category, These correspond to high stress hazard, medium stress hazard, and low stress hazard, respectively. Finally, normalization is performed to obtain the output result: 1) Indicators: Pressure signal increases sharply or fluctuates violently, drill cuttings flow rate far exceeds the normal amount and the particle size is fine, and the maximum energy of micro-vibration is >10. 5 J; Hazard level: High stress hazard; 2) Indicators: Increased or drastic pressure signal, increased drill cuttings flow rate and smaller particle size, and maximum micro-vibration energy of 10. 2 -10 5 J; Hazard level: Moderate stress hazard; 3) Indicators: Pressure signal fluctuates slightly; drill cuttings flow rate and particle size variation range is within 5% above and below normal drilling; maximum micro-vibration energy is 10. 2 -10 4 J; Hazard level: Weak stress hazard.
8. The method for real-time monitoring and early warning of surrounding rock stress during drilling as described in claim 7, characterized in that, In step S5, the update process of the core parameters of the surrounding rock stress monitoring model is as follows: In the loss calculation, the true class label of each sample is y, which is converted into a one-hot encoding Y. The loss function is expressed as: ； in, Is the i-th sample in the true category? The predicted probability; It is the average loss of N samples in the same batch; The gradient of the loss function with respect to the unnormalized class score z of the Softmax input: ； in, Is the i-th sample in the true category? The predicted probability; N is the number of samples in the same batch; In backpropagation, the loss is applied to the weights and bias vectors as well as the LSTM hidden vectors. The gradient is expressed as: ； ； in It is the first The unnormalized score vector of each sample output; ； Then use gradient descent to update the parameters; Weight update: ； in, The updated weights; Bias vector update: ； in, The learning rate in gradient descent. This is the updated bias vector.