A Smart Early Warning Method for Surrounding Rock Stability in Deep Coal Seam Roadways

Abstract

This invention provides an intelligent early warning method for the stability of surrounding rock in deep coal seam roadways, relating to the field of coal mine dynamic disaster prevention technology. The method constructs a basic sample library through dynamic-static combined loading rock tests; simultaneously collects and time-series aggregates static stress and microseismic data of the surrounding rock; constructs an equivalent dynamic-static combined stress index based on the principle of dynamic-static load superposition, forming a seismic-force coupled feature vector; establishes an LSTM early warning model, employs a transfer learning strategy, pre-trains using laboratory data, and then fine-tunes with field data; inputs the real-time feature vector into the model to output the instability probability; verifies and corrects the probability based on actual surrounding rock deformation data, and uses missed samples for online incremental model updates. This invention quantifies the dynamic-static load coupling effect, overcomes the difficulty of small sample training, and achieves accurate and adaptive early warning of the instability risk of surrounding rock in deep roadways through a physical feedback closed-loop correction mechanism.

Description

Technical Field

[0001] This invention relates to the field of coal mine dynamic disaster prevention and control technology, specifically to an intelligent early warning method for the stability of surrounding rock in deep coal seam roadways based on the spatiotemporal evolution characteristics of microseismic stress. Background Technology

[0002] With the increasing depth of coal mining, the geomechanical environment faced by deep coal seam roadways is becoming increasingly complex, exhibiting significant characteristics of "high ground stress" and "strong mining disturbance." Under this high-stress environment, the stability of the surrounding rock in the roadway is affected not only by static geological conditions but also by dynamic load disturbances induced by dynamic events such as roof fractures. This nonlinear superposition of static and dynamic loads is the main cause of dynamic disasters such as rockbursts. Currently, using deep learning and multiphysics monitoring technologies for early warning of dynamic disasters such as rockbursts has become a research hotspot, but existing technologies still have the following shortcomings:

[0003] First, existing methods often treat microseismic and stress data as independent input channels, neglecting the nonlinear disaster-causing mechanism of dynamic and static load superposition, making it difficult to quantify the weakening effect of dynamic load on the static load bearing limit. Second, the scarcity of samples of catastrophic events such as rockbursts makes it difficult to train pure data-driven models and results in poor generalization ability. Third, existing early warning systems lack a closed-loop correction mechanism based on physical deformation feedback, making it impossible to verify prediction results in real time and update the model online, leading to a decrease in robustness over time.

[0004] For example, Chinese patent document CN117312919A discloses a rockburst early warning method based on a BO-CNN-LSTM model, but it is a purely data-driven model lacking physical mechanism constraints and does not consider the dynamic-static load coupling effect. Chinese patent document CN111599137A discloses a multi-physics field monitoring and early warning system and method for the stability of surrounding rock in underground engineering. Although it integrates multi-physics field data, the feature fusion method is simple and lacks a dedicated physical index construction and closed-loop correction mechanism. Summary of the Invention

[0005] This invention provides an intelligent early warning method for the stability of surrounding rock in deep coal seam roadways. By constructing a feature vector that reflects the superposition effect of dynamic and static loads, and combining transfer learning and LSTM network, it can accurately identify the impact risk of surrounding rock in deep coal seam roadways, and introduce physical deformation feedback for closed-loop correction.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for intelligent early warning of surrounding rock stability in deep coal seam roadways, the method comprising the following steps:

[0008] S1. Construct a basic sample library containing input feature vectors and stability labels through dynamic and static combined loading rock tests;

[0009] S2. Simultaneously acquire static stress and microseismic data of the surrounding rock at the site, and map them to a continuous time window for time-series aggregation;

[0010] S3. Standardize the aggregated data and construct a seismic-force coupling feature vector based on the principle of dynamic and static load superposition.

[0011] S4. Establish an early warning model based on long short-term memory networks;

[0012] S5. Adopting a transfer learning strategy, first pre-train the model based on the S1 basic sample library, then fine-tune it using the on-site collected data in S2, and validate it using the test set.

[0013] S6. Input the real-time collected and processed seismic-force coupling feature vector into the model and output the instability probability.

[0014] S7. The instability probability is verified and corrected based on the actual deformation data of the surrounding rock, and the missed samples are added to the difficult case sample library to update the model online.

[0015] Further, S1 includes:

[0016] S11. Drill hard rock and soft rock samples in the surrounding rock of deep coal seam roadways;

[0017] S12. Process the drilled rock samples into rock specimens of standard size;

[0018] S13. A rock triaxial compression loading tester and a pendulum impact tester were used for combined loading, applying static axial compression simulating the original rock stress and dynamic impact simulating vibration, respectively.

[0019] S14. Record the rock failure state under different combinations of working conditions and construct a basic sample library. :

[0020] ;

[0021] in,

[0022] The input features for the i-th test include the applied static axial stress value and the dynamic impact energy value;

[0023] For stability labels, 0 is marked if the rock specimen remains intact, and 1 is marked if the rock specimen is damaged.

[0024] K represents the total number of test samples, which is the sum of the effective number of tests under different dynamic and static combined loading conditions.

[0025] Further, S2 includes:

[0026] S21. Set up monitoring stations at intervals in key monitoring areas of deep roadways. Each monitoring station is equipped with a deep-hole multi-point displacement gauge, an anchor cable force gauge and a micro-vibration pickup array.

[0027] S22. Real-time acquisition of anchor bolt / cable force sequence σ(t), and by capturing elastic wave signals generated by micro fractures in the rock mass through a microseismic sensor array, the energy sequence E(t) of the microseismic event is calculated.

[0028] S23. Map discrete microseismic data to a continuous time window. Above, define the first Polymer stress within a time window and aggregate energy for:

[0029] ;

[0030] ;

[0031] in,

[0032] For time window The number of stress sampling points within; For time window The total number of microseismic events that occurred within the area.

[0033] Further, S3 includes:

[0034] S31. Clean the collected data, fill in missing data and remove outliers and noise.

[0035] S32. Standardize each physical quantity using the Z-Score method. The standardized values ​​are... :

[0036] ;

[0037] in,

[0038] This is the original monitoring data; The mean of the feature data; Standard deviation;

[0039] S33. Construct an equivalent dynamic-static combined stress index with lithology self-adaptive capability. The calculation formula is as follows:

[0040] ;

[0041] in,

[0042] The static stress aggregation value of the surrounding rock monitored at time t; E0 represents the aggregated energy released by the microseismic event at time t; E0 is the energy normalization baseline value. is the dynamic-static load dimension conversion coefficient, whose value is positively correlated with the impact tendency index of the surrounding rock of the roadway; ln is the natural logarithm function;

[0043] S34. First-order difference characteristics of the calculation index :

[0044] ;

[0045] in,

[0046] The equivalent dynamic and static combined stress index aggregate value after standardization within the k-th time window; The equivalent dynamic and static combined stress index aggregate value after standardization within the (k-1)th time window; To represent the time interval between two consecutive time windows;

[0047] S35. Constructing the model input feature vector :

[0048] ;

[0049] in,

[0050] This represents data that has undergone standardization. The normalized static stress aggregate value of the surrounding rock at time k; The standardized aggregate energy value of the microseismic event at time k; The value of the equivalent dynamic-static combined stress index at time k is the standardized value.

[0051] Furthermore, the long short-term memory network early warning model established in S4, for time... input vector The state is updated using a forget gate, an input gate, and an output gate:

[0052] Computational forget gate:

[0053] ;

[0054] in,

[0055] The output vector of the forget gate has a value between (0,1), representing the proportion of cell state information retained from the previous time step. The closer the value is to 1, the more information is retained; the closer it is to 0, the more information is forgotten. Sigmoid is the activation function. Here is the weight matrix for the forget gate; This represents the hidden layer state from the previous time step. This is the input feature vector at the current time. This is the bias vector for the forget gate;

[0056] Calculate the input gate:

[0057] ;

[0058] ;

[0059] in,

[0060] The output vector of the input gate determines what new information will be updated to the cell state at the current moment; The candidate cell state at the current moment is a new vector containing new information created through the tanh layer; tanh is the hyperbolic tangent activation function. and The weight matrices are calculated for the input gate and the candidate state, respectively. and These are the corresponding bias vectors;

[0061] Perform a status update:

[0062] = ⊙ + ⊙ ;

[0063] in,

[0064] The cell state at the current moment is calculated by forgetting old information and adding new candidate information; The cell state at the previous time step; ⊙ represents the Hadamard product, which is the operation of multiplying corresponding elements of a matrix or vector;

[0065] Output the hidden layer features and hidden states at the current time step:

[0066] ;

[0067] = ⊙ ;

[0068] in,

[0069] The output vector of the output gate determines which information in the current cell state will be output. The closer its value is to 1, the more information will be output, and the closer it is to 0, the more information will be masked. This is the hidden state at the current time step, which is the final feature representation after filtering by the output gate. It will be passed to the LSTM unit at the next time step. The weight matrix of the output gate is used to perform a weighted summation of the input data and the hidden state at the previous time step, determining the emphasis of the output information; This is the bias term for the output gate, used to adjust the threshold of the activation function and improve the model's fitting ability.

[0070] Calculate the output layer probability:

[0071] ;

[0072] in,

[0073] The value is the probability value of the surrounding rock becoming unstable at the current time t, which is predicted by the model. Its value ranges between [0,1]. The closer the value is to 1, the higher the risk of instability. The closer the value is to 0, the more stable the surrounding rock is. The weight matrix of the output layer is used to map the hidden feature vectors output by the LSTM to a one-dimensional decision space. Here, represents the bias term for the output layer; e is the natural constant, the base of the natural logarithm; furthermore, the model training process employs a binary cross-entropy loss function with an added L2 regularization term. The loss function is:

[0074] ;

[0075] in,

[0076] This represents the batch sample size. This is a real label; Predict probabilities for the model; For L2 regularization terms, The regularization coefficient is used. These are all network weight parameters.

[0077] Further, S5 includes:

[0078] S51. Use rock test data with dynamic and static loading to pre-train the model. Similarity ratio conversion is required before training.

[0079] S52. Adopt a transfer learning strategy, keep the weight parameters of the feature extraction layer in the pre-trained model unchanged, and only unfreeze and fine-tune the parameters of the classification decision layer of the model.

[0080] S53. Calculate the F1 score of the model on the test set. If the F1 score is lower than the set threshold, adjust the hyperparameters and return to training.

[0081] Further, S6 includes:

[0082] S61. Generate a real-time seismic-force coupling feature vector by combining the real-time acquired static stress of the surrounding rock with the microseismic data. Input the early warning model;

[0083] S62. The model outputs the instability probability value P of the surrounding rock at the current moment.

[0084] Furthermore, based on the instability probability value P, a three-level warning system is set: P < 0.5 is the green safe zone, 0.5 ≤ P < 0.8 is the yellow warning zone, and P ≥ 0.8 is the red alarm zone.

[0085] Further, S7 includes:

[0086] S71. The actual deformation of the surrounding rock is obtained in real time through multi-point displacement gauges on site. Compare with the set limit displacement threshold Calculate the normalized deformation ratio :

[0087] ;

[0088] S72. Constructing the revised final early warning probability :

[0089] ;

[0090] in,

[0091] Confidence weights; The original predicted probabilities of the model; To correct the linear unit;

[0092] S73, Online Model Evolution:

[0093] Define when and If a data point is missed, it is considered a "missed report," and the data for that period is added to the difficult case sample library for incremental training.

[0094] in,

[0095] This is the rate threshold.

[0096] Compared with existing technologies, the intelligent early warning method for the stability of surrounding rock in deep coal seam roadways of the present invention achieves the following beneficial technical effects:

[0097] 1. This invention achieves the scientific quantification of the dynamic-static load coupling effect by constructing an "equivalent dynamic-static combined stress" index. Traditional early warning methods often process microseismic and stress data independently, failing to effectively reflect the physical essence of the disaster caused by the superposition of dynamic and static loads. This invention couples the static stress collected on-site with the aggregated value of microseismic energy to construct a feature vector characterizing the synergistic effect of dynamic and static loads. This elevates the disaster-causing mechanism from isolated monitoring of multi-source information to a direct quantitative characterization of the nonlinear superposition process of "static load bearing capacity" and "dynamic load disturbance weakening." This fundamentally overcomes the shortcomings of existing technologies in characterizing the weakening effect of dynamic loads on the bearing capacity limit of surrounding rock, giving the early warning model clear physical mechanism constraints and significantly improving the scientific nature and accuracy of early warning criteria.

[0098] 2. This invention innovatively employs a hybrid modeling strategy of "laboratory pre-training + field data transfer learning," effectively overcoming the model training dilemma caused by the scarcity of disaster samples. Addressing the problem of scarce samples from disaster events such as rockbursts and the poor generalization ability of purely data-driven models, this invention first constructs a basic sample library in the laboratory through dynamic and static combined loading experiments, allowing the model to learn the basic damage evolution laws of rocks under force-seismic coupling. Subsequently, using transfer learning technology, the knowledge of the laboratory pre-trained model is transferred to the engineering field, requiring only minor adjustments with a small amount of field data to adapt to specific geological environments. This strategy significantly reduces the dependence on massive amounts of field disaster samples, ensuring the model's rapid convergence and stable generalization ability under small sample conditions, thus solving the core bottleneck of data-driven model application in engineering practice.

[0099] 3. This invention establishes a closed-loop correction and online update mechanism based on physical deformation feedback, significantly improving the system's long-term robustness and adaptability. Existing early warning systems are mostly open-loop, with models fixed once deployed, unable to adapt to dynamic changes in geological conditions. This invention introduces real-time surrounding rock deformation monitoring data as a physical verification benchmark, applying confidence-weighted correction to the model's output instability probability. This uses physical measurements to constrain model prediction bias, effectively reducing false alarms and missed alarms. More importantly, the system can automatically identify data from "missed alarm" periods as difficult examples, enabling online incremental learning of the model and achieving continuous self-optimization and evolution of the early warning model during operation. This allows the system to dynamically adapt to the impact of mining disturbances and changes in geological conditions, ensuring the long-term stability of early warning accuracy.

[0100] In summary, this invention organically combines three core technologies—mechanism-driven feature fusion, small-sample adaptive hybrid modeling, and closed-loop evolution based on physical feedback—to form a complete early warning technology system encompassing data perception, intelligent identification, and dynamic optimization. This not only enables more accurate and timely intelligent perception and early warning of the risk of rock instability in deep roadways but also fundamentally enhances the adaptability, reliability, and practicality of the early warning system under complex geological conditions, providing strong technical support for the safe and efficient mining of deep coal mines. Attached Figure Description

[0101] Figure 1 This is a schematic diagram of the location for core sampling of rock samples in the surrounding rock of a deep coal seam roadway, according to an embodiment of the present invention.

[0102] Figure 2 The hard rock specimen processed according to an embodiment of the present invention;

[0103] Figure 3 The soft rock specimen processed according to an embodiment of the present invention;

[0104] Figure 4 This is a schematic diagram of a rock triaxial compression loading test machine and a pendulum impact test machine according to an embodiment of the present invention;

[0105] Figure 5 This is a plan view of the layout of the measuring stations according to an embodiment of the present invention;

[0106] Figure 6 This is an architecture diagram of the early warning model according to an embodiment of the present invention;

[0107] Figure 7 This is a flowchart of a method according to an embodiment of the present invention;

[0108] Figure 8 This is a schematic diagram of the online feedback correction and closed-loop optimization logic of the model in an embodiment of the present invention. Detailed Implementation

[0109] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. Certain embodiments of the invention will be described more fully below with reference to the accompanying drawings, and some, but not all, of these embodiments will be shown. In fact, various embodiments of the invention can be implemented in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided to enable the invention to meet applicable legal requirements.

[0110] In the description of this invention, it should be noted that the terms "inner," "outer," "upper," "lower," "front," and "rear," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0111] This embodiment is based on the actual situation of a coal mine and refers to... Figures 1 to 8 As shown, an intelligent early warning method for the stability of surrounding rock in deep coal seam roadways is provided. The method includes the following steps:

[0112] S1. Construct a basic sample library containing input feature vectors and stability labels through dynamic and static combined loading rock tests.

[0113] S11, such as Figure 1 As shown, hard rock and soft rock samples were drilled in the surrounding rock of the deep coal seam roadway. Three core sampling points were selected on site, and three rock core samples were drilled from the two sides and the roof at each core sampling point. A total of nine rock core samples were drilled on site, and the length of each core sampling hole was 15m.

[0114] S12. Process the drilled rock samples into standard-sized rock specimens.

[0115] The rock core samples drilled on-site were wrapped in plastic wrap to prevent contact with air and water, and then transported to a processing plant for standardized processing to prepare standard-sized (50mm × 100mm) rock specimens. Figure 2 , Figure 3 As shown, a total of 15 hard rock specimens and 15 soft rock specimens were prepared.

[0116] S13, such as Figure 4 As shown, a rock triaxial compression loading tester and a pendulum impact testing machine were used for combined loading, applying static axial compression at different gradients. Simulating original rock stress and dynamic impact Simulated vibration.

[0117] S14. Record the rock failure state under different combinations of working conditions and construct a basic sample library. :

[0118] ;

[0119] in,

[0120] The input features for the i-th test include the applied static axial stress value and the dynamic impact energy value;

[0121] For stability labels, 0 is marked if the rock specimen remains intact, and 1 is marked if the rock specimen is damaged.

[0122] K represents the total number of test samples, which is the sum of the effective number of tests under different dynamic and static combined loading conditions. In this test, K is set to 30.

[0123] S2. Synchronously acquire static stress and microseismic data of the surrounding rock at the site, and map them to a continuous time window for time-series aggregation.

[0124] S21. Measuring stations are deployed at intervals in key monitoring areas of deep roadways. Each station is equipped with a deep-hole multi-point displacement gauge, a bolt cable force gauge (sampling frequency set to 1-10Hz), and a micro-seismic pickup array. After installation, relevant software is used for debugging. Specifically, three sets of measuring stations are deployed at 100m intervals on-site.

[0125] S22. The anchor / cable stress sequence σ(t) is acquired in real time, and the elastic wave signal generated by the microfracture of the rock mass is captured by a microseismic sensor array to calculate the energy sequence E(t) of the microseismic event. A slow increase in σ(t) indicates that the surrounding rock is undergoing compression deformation and the static load is increasing. A sudden and violent fluctuation in σ(t) indicates that the internal structure of the rock mass has been adjusted. E(t) records the amount of energy released by the microseismic event occurring at time t.

[0126] S23. Map discrete microseismic data to a continuous time window. Above, define the first Polymer stress within a time window and aggregate energy for:

[0127] ;

[0128] ;

[0129] in,

[0130] For time window The number of stress sampling points within; For time window The total number of microseismic events that occurred within the area.

[0131] S3. Standardize the aggregated data and construct the seismic-force coupling feature vector based on the principle of dynamic and static load superposition.

[0132] S31. Clean the collected data, use Lagrange interpolation to complete missing data and remove outliers.

[0133] S32. Standardize each physical quantity using the Z-Score method. The standardized values ​​are... :

[0134] ;

[0135] in,

[0136] This is the original monitoring data; The mean of the feature data; The standard deviation is denoted as .

[0137] S33. Construct an equivalent dynamic-static combined stress index with lithology self-adaptive capability. The calculation formula is as follows:

[0138] ;

[0139] in,

[0140] The static stress aggregation value of the surrounding rock monitored at time t; E0 represents the aggregated energy released by the microseismic event at time t; E0 is the energy normalization baseline value. is the dynamic-static load dimension conversion coefficient, whose value is positively correlated with the impact tendency index of the surrounding rock of the roadway; ln is the natural logarithm function.

[0141] S34. To capture the abrupt change trends before a disaster, calculate the first-order difference characteristics of the indicators. :

[0142] ;

[0143] in,

[0144] The equivalent dynamic and static combined stress index aggregate value after standardization within the k-th time window; This is the aggregated value of the equivalent dynamic and static combined stress index after standardization within the (k-1)th time window (i.e., the previous moment); This represents the time interval (sampling step size) between two consecutive time windows.

[0145] S35. Constructing the model input feature vector :

[0146] ;

[0147] in,

[0148] This represents data that has undergone standardization. The normalized static stress aggregate value of the surrounding rock at time k; The standardized aggregate energy value of the microseismic event at time k; The value of the equivalent dynamic-static combined stress index at time k is the standardized value.

[0149] S4. Establish an early warning model based on long short-term memory networks and define the network structure to explore the temporal evolution of feature vectors.

[0150] For time input vector The state is updated using a forget gate, an input gate, and an output gate:

[0151] Computational forget gate:

[0152] ;

[0153] in,

[0154] The output vector of the forget gate has a value between (0,1), representing the proportion of cell state information retained from the previous time step. The closer the value is to 1, the more information is retained; the closer it is to 0, the more information is forgotten. Sigmoid is the activation function. Here is the weight matrix for the forget gate; This represents the hidden layer state from the previous time step. This is the input feature vector at the current time. This is the bias vector for the forget gate;

[0155] Calculate the input gate:

[0156] ;

[0157] ;

[0158] in,

[0159] The output vector of the input gate determines what new information will be updated to the cell state at the current moment; The candidate cell state at the current moment is a new vector containing new information created through the tanh layer; tanh is the hyperbolic tangent activation function. and The weight matrices are calculated for the input gate and the candidate state, respectively. and These are the corresponding bias vectors;

[0160] Perform a status update:

[0161] = ⊙ + ⊙ ;

[0162] in,

[0163] The cell state at the current moment is calculated by forgetting old information and adding new candidate information; The cell state at the previous time step; ⊙ represents the Hadamard product, which is the operation of multiplying corresponding elements of a matrix or vector.

[0164] Output the hidden layer features and hidden states at the current time step:

[0165] ;

[0166] = ⊙ ;

[0167] in,

[0168] The output vector of the output gate determines which information in the current cell state will be output. The closer its value is to 1, the more information will be output, and the closer it is to 0, the more information will be masked. This is the hidden state at the current time step, which is the final feature representation after filtering by the output gate. It will be passed to the LSTM unit at the next time step. The weight matrix of the output gate is used to perform a weighted summation of the input data and the hidden state at the previous time step, determining the emphasis of the output information; This is the bias term for the output gate, used to adjust the threshold of the activation function and improve the model's fitting ability.

[0169] Calculate the output layer probability:

[0170] ;

[0171] in,

[0172] The value is the probability value of the surrounding rock becoming unstable at the current time t, which is predicted by the model. Its value ranges between [0,1]. The closer the value is to 1, the higher the risk of instability. The closer the value is to 0, the more stable the surrounding rock is. The weight matrix of the output layer is used to map the hidden feature vectors output by the LSTM to a one-dimensional decision space. Here, represents the bias term for the output layer; e is the natural constant, the base of the natural logarithm; the model training process uses a binary cross-entropy loss function with added L2 regularization. The loss function is:

[0173] ;

[0174] in,

[0175] This represents the batch sample size. This is a real label; Predict probabilities for the model; For L2 regularization terms, The regularization coefficient is used. These are all network weight parameters.

[0176] S5. Using a transfer learning strategy, the model is first pre-trained based on the S1 basic sample library to learn the basic temporal characteristics of rocks under force-seismic coupling. Then, the model is fine-tuned using the S2 field-collected data and validated using the test set.

[0177] S51. The model is pre-trained using rock test data with dynamic and static loading. Since the size of the rock specimen (50mm×100mm) is very different from the scale of the roadway on site, and the loading energy level in the laboratory is low, a similarity ratio conversion is required before training so that the model can learn the basic "stress-vibration-damage" evolution law.

[0178] S52. A transfer learning strategy is adopted to keep the weight parameters of the feature extraction layer in the pre-trained model unchanged. This layer has learned the basic temporal characteristics of rocks under force-seismic coupling in the dynamic and static combined loading rock test data. Only the parameters of the classification decision layer of the model are unfrozen and fine-tuned. The parameters of this layer are updated using a small amount of field-collected data to adapt to the specific geological environment and noise distribution on site.

[0179] S53. Calculate the F1 score of the model on the test set. If the F1 score is lower than the set threshold (0.8), adjust the hyperparameters. They then returned to training.

[0180] ;

[0181] in,

[0182] TP (True Positive) means that the model "successfully captured the danger of surrounding rock instability"; FP (False Positive) means that the model issued a "false warning" when the surrounding rock was stable; FN (False Negative) means that the model did not issue any warning for surrounding rock instability.

[0183] S6. Input the real-time collected and processed seismic-force coupling feature vector into the model, and output the instability probability and early warning level.

[0184] S61. Generate a real-time seismic-force coupling feature vector by combining the real-time acquired static stress of the surrounding rock with the microseismic data. Input the early warning model.

[0185] S62. The model outputs the instability probability value P of the surrounding rock at the current moment.

[0186] Based on the instability probability value P, three warning levels are set: when P < 0.5, it is a green safety zone, the surrounding rock is in a stable state, and normal monitoring is maintained; when 0.5 ≤ P < 0.8, it is a yellow warning zone, the monitoring indicators show abnormal fluctuations, indicating that support needs to be strengthened or the tunneling speed needs to be reduced; when P ≥ 0.8, it is a red alarm zone, the surrounding rock is critically unstable, the system automatically triggers an audible and visual alarm, and notifies on-site personnel to evacuate immediately through the terminal.

[0187] S7. The instability probability is verified and corrected based on the actual deformation data of the surrounding rock, and the missed samples are added to the difficult case sample library to update the model online.

[0188] Physical verification: The actual deformation of the surrounding rock is acquired in real time, and the physical deformation deviation index is calculated by comparing the actual deformation with the limit displacement threshold.

[0189] Probability Correction: The initial predicted probability of the output is weighted and fused to correct the deviation of the physical deformation index, and the final warning level is generated. The model prediction deviation is constrained by the physical measured data.

[0190] Online update: Real-time monitoring of surrounding rock deformation rate; identifying missed cases where the model predicts the rock to be safe but the actual deformation rate exceeds a preset threshold; adding these cases to the training library for online incremental updates to the model, thus enabling the adaptive evolution of the early warning model.

[0191] S71. The actual deformation of the surrounding rock is obtained in real time through multi-point displacement gauges on site. Compare with the set limit displacement threshold Calculate the normalized deformation ratio :

[0192] .

[0193] S72. Constructing the revised final early warning probability :

[0194] ;

[0195] in,

[0196] Confidence weights; The original predicted probabilities of the model; To correct the linear unit.

[0197] S73, Online Model Evolution:

[0198] Define when and If a data point is missed, it is marked as y=1 and added to the hard case sample library for incremental training.

[0199] in,

[0200] This is the rate threshold.

[0201] The present invention has been described in detail above with reference to the accompanying drawings. Based on the above description, those skilled in the art should have a clear understanding of the intelligent early warning method for the stability of surrounding rock in deep coal seam roadways according to the present invention. The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent early warning of surrounding rock stability in deep coal seam roadways, characterized in that, The method includes the following steps: S1. Construct a basic sample library containing input feature vectors and stability labels through dynamic and static combined loading rock tests; S2. Simultaneously acquire static stress and microseismic data of the surrounding rock at the site, and map them to a continuous time window for time-series aggregation; S3. Standardize the aggregated data and construct a seismic-force coupling feature vector based on the principle of dynamic and static load superposition. S3 includes: S31. Clean the collected data, fill in missing data and remove outliers and noise. S32. Standardize each physical quantity using the Z-Score method. The standardized values ​​are... : ； in, This is the original monitoring data; The mean of the feature data; Standard deviation; S33. Construct an equivalent dynamic-static combined stress index with lithology self-adaptive capability. The calculation formula is as follows: ； in, The static stress aggregation value of the surrounding rock monitored at time t; E0 represents the aggregated energy released by the microseismic event at time t; E0 is the energy normalization baseline value. is the dynamic-static load dimension conversion coefficient, whose value is positively correlated with the impact tendency index of the surrounding rock of the roadway; ln is the natural logarithm function; S34. First-order difference characteristics of the calculation index : ； in, The equivalent dynamic and static combined stress index aggregate value after standardization within the k-th time window; The equivalent dynamic and static combined stress index aggregate value after standardization within the (k-1)th time window; The time interval between two consecutive time windows; S35. Constructing the model input feature vector : ； in, This represents data that has undergone standardization. The standardized static stress aggregate value of the surrounding rock at time k; The standardized aggregate energy value of the microseismic event at time k; The equivalent dynamic-static combined stress index value at time k after standardization; S4. Establish an early warning model based on long short-term memory networks; The long short-term memory network early warning model established in S4, for time... input vector The state is updated using a forget gate, an input gate, and an output gate: Computational forget gate: ； in, The output vector of the forget gate has a value between (0,1), representing the proportion of cell state information retained from the previous time step. The closer the value is to 1, the more information is retained; the closer it is to 0, the more information is forgotten. Sigmoid is the sigmoid activation function. Here is the weight matrix for the forget gate; This represents the hidden layer state from the previous time step. This is the input feature vector at the current time. This is the bias vector for the forget gate; Calculate the input gate: ； ； in, The output vector of the input gate determines what new information will be updated to the cell state at the current moment; The candidate cell state at the current moment is a new vector containing new information created through the tanh layer; tanh is the hyperbolic tangent activation function. and The weight matrices are calculated for the input gate and the candidate state, respectively. and These are the corresponding bias vectors; Perform a status update: = ⊙ + ⊙ ； in, The cell state at the current moment is calculated by forgetting old information and adding new candidate information; The cell state at the previous time step; ⊙ represents the Hadamard product, which is the operation of multiplying corresponding elements of a matrix or vector; Output the hidden layer features and hidden states at the current time step: ； = ⊙ ； in, The output vector of the output gate determines which information in the current cell state will be output. The closer its value is to 1, the more information will be output, and the closer it is to 0, the more information will be masked. This is the hidden state at the current time step, which is the final feature representation after filtering by the output gate. It will be passed to the LSTM unit at the next time step. The weight matrix of the output gate is used to perform a weighted summation of the input data and the hidden state at the previous time step, determining the emphasis of the output information; This is the bias term for the output gate, used to adjust the threshold of the activation function and improve the model's fitting ability. Calculate the output layer probability: ； in, The value is the probability value of the surrounding rock becoming unstable at the current time t, which is predicted by the model. Its value ranges between [0,1]. The closer the value is to 1, the higher the risk of instability. The closer the value is to 0, the more stable the surrounding rock is. The weight matrix of the output layer is used to map the hidden feature vectors output by the LSTM to a one-dimensional decision space. is the bias term for the output layer; e is the natural constant, the base of the natural logarithm; S5. Adopting a transfer learning strategy, first pre-train the model based on the S1 basic sample library, then fine-tune it using the on-site collected data in S2, and validate it using the test set. S6. Input the real-time collected and processed seismic-force coupling feature vector into the model and output the instability probability. S7. The instability probability is verified and corrected based on the actual deformation data of the surrounding rock, and the missed samples are added to the difficult case sample library to update the model online.

2. The intelligent early warning method for the stability of surrounding rock in deep coal seam roadways according to claim 1, characterized in that, S1 includes: S11. Drill hard rock and soft rock samples in the surrounding rock of deep coal seam roadways; S12. Process the drilled rock samples into rock specimens of standard size; S13. A rock triaxial compression loading tester and a pendulum impact tester were used for combined loading, applying static axial compression simulating the original rock stress and dynamic impact simulating vibration, respectively. S14. Record the rock failure state under different combinations of working conditions and construct a basic sample library. : ； in, The input features for the i-th test include the applied static axial stress value and the dynamic impact energy value; For stability labels, 0 is marked if the rock specimen remains intact, and 1 is marked if the rock specimen is damaged. K represents the total number of test samples, which is the sum of the effective number of tests under different dynamic and static combined loading conditions.

3. The intelligent early warning method for the stability of surrounding rock in deep coal seam roadways according to claim 1, characterized in that, S2 includes: S21. Set up monitoring stations at intervals in key monitoring areas of deep roadways. Each monitoring station is equipped with a deep-hole multi-point displacement gauge, an anchor cable force gauge and a micro-vibration pickup array. S22. Real-time acquisition of anchor bolt / cable force sequence σ(t), and by capturing the elastic wave signal generated by micro fractures in the rock mass through a microseismic sensor array, the energy sequence E(t) of the microseismic event is calculated. S23. Map discrete microseismic data to a continuous time window. Above, define the first Polymer stress within a time window and aggregate energy for: ； ； in, For time window The number of stress sampling points within; For time window The total number of microseismic events that occurred within the area.

4. The intelligent early warning method for the stability of surrounding rock in deep coal seam roadways according to claim 1, characterized in that, The model training process uses a binary cross-entropy loss function with an added L2 regularization term. The loss function is: ； in, This represents the batch sample size. This is a real label; Predict probabilities for the model; For L2 regularization terms, The regularization coefficient is used. These are all network weight parameters.

5. The intelligent early warning method for the stability of surrounding rock in deep coal seam roadways according to claim 1, characterized in that, S5 includes: S51. Use rock test data with dynamic and static loading to pre-train the model. Similarity ratio conversion is required before training. S52. Adopt a transfer learning strategy, keep the weight parameters of the feature extraction layer in the pre-trained model unchanged, and only unfreeze and fine-tune the parameters of the classification decision layer of the model. S53. Calculate the F1 score of the model on the test set. If the F1 score is lower than the set threshold, adjust the hyperparameters and return to training.

6. The intelligent early warning method for the stability of surrounding rock in deep coal seam roadways according to claim 1, characterized in that, S6 includes: S61. Generate a real-time seismic-force coupling feature vector by combining the real-time acquired static stress of the surrounding rock with the microseismic data. Input the early warning model; S62. The model outputs the instability probability value P of the surrounding rock at the current moment.

7. The intelligent early warning method for the stability of surrounding rock in deep coal seam roadways according to claim 6, characterized in that, Based on the instability probability value P, three warning levels are set: green safety zone when P < 0.5, yellow warning zone when 0.5 ≤ P < 0.8, and red alarm zone when P ≥ 0.

8.

8. The intelligent early warning method for the stability of surrounding rock in deep coal seam roadways according to claim 1, characterized in that, S7 includes: S71. The actual deformation of the surrounding rock is obtained in real time through multi-point displacement gauges on site. Compare with the set limit displacement threshold Calculate the normalized deformation ratio : ； S72. Constructing the revised final early warning probability : ； in, Confidence weights; The original predicted probabilities of the model; To correct the linear unit; S73, Online Model Evolution: Define when and If a data point is missed, it is considered a "missed report," and the data for that period is added to the difficult case sample library for incremental training. in, This is the rate threshold.

Images