Mining rock mass structure degradation monitoring method and system based on deep learning

Abstract

The invention discloses a mining rock mass structure degradation monitoring method and system based on deep learning, particularly relates to the technical field of mine safety monitoring, and is used for solving the problems of monitoring lag and misjudgment caused by the fact that an existing static model cannot automatically adapt to rock mass damage dynamic evolution. The method comprises the following steps: constructing a dynamic damage incremental data set by collecting multi-source heterogeneous monitoring data under mining disturbance in real time, extracting newly-added damage features, and calculating a distribution offset degree between the newly-added damage features and historical features; when the deviation degree exceeds the limit, verifying and screening a damage feature subset according with a mechanical law based on a rock constitutive equation, and blocking non-physical noise interference; a geological structure evolution mode is matched through damage path dependence modeling, and a potential damage path thermodynamic diagram is generated in combination with a space gradient; performing cross-stage association on the thermodynamic diagram and a historical feature library by adopting a knowledge distillation mechanism to generate a dynamic weight matrix fused with a mining time sequence; and finally, updating the damage assessment model through a parameter reweighting mechanism to realize adaptive assessment of the rock mass degradation risk.

Description

Technical Field

[0001] The present invention relates to the technical field of mine safety monitoring. More specifically, the present invention relates to a method and system for monitoring the deterioration of mined rock mass structures based on deep learning. Background Art

[0002] In the field of mine safety monitoring, data-driven methods for evaluating the deterioration of rock mass structures have gradually replaced traditional manual observation methods. Existing technologies generally use acoustic, seismic, and optical data of rock masses collected during historical mining stages to construct a training set, and establish a mapping relationship between rock mass damage characteristics and stability through machine learning models, thereby realizing automated monitoring and early warning. Relying on the data distribution assumption during the training stage, the trained model is directly deployed in monitoring scenarios at different mining times.

[0003] Due to the irreversible and dynamic development characteristics of rock mass damage evolution, there are essential differences between the mechanical behaviors of newly generated fractures and historical damage patterns during the mining process, resulting in the inability of the static training set to cover the new feature space of real-time monitoring data. Existing methods lack the ability to autonomously adapt to the dynamic deterioration law of rock mass structures after model deployment, causing the evaluation results of the monitoring model under new damage patterns to deviate from the true physical state, forming a monitoring blind area, which directly affects the accuracy and timeliness of mine safety early warning. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a method and system for monitoring the deterioration of mined rock mass structures based on deep learning to solve the problems raised in the above background art.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for monitoring the deterioration of mined rock mass structures based on deep learning, comprising the following steps:

[0007] S1. Real-time collect multi-source heterogeneous monitoring data under mining disturbances, and construct a dynamic damage increment data set;

[0008] S2. Extract the newly added damage characteristics of the current mining stage based on the dynamic damage increment data set, and generate a dynamic distribution offset degree based on the similarity measurement between the newly added damage characteristics and the historical damage characteristic library;

[0009] S3. When the dynamic distribution offset degree exceeds the offset degree threshold, perform mechanical behavior constraint verification on the newly added damage characteristics, and screen out an effective damage characteristic subset that conforms to the mechanical law of rock mass failure;

[0010] S4. Perform damage path dependence modeling on the effective damage characteristic subset, match the damage evolution pattern in the historical damage characteristic library in the same geological structure area, and generate a potential failure path heat map based on the spatial gradient;

[0011] S5. Adopt a knowledge distillation mechanism to conduct cross-stage correlation analysis on the potential damage path heat map and the historical damage feature library, and generate a dynamic weight matrix reflecting the mapping relationship between the mining time sequence and damage evolution;

[0012] S6. Re-weight the parameters of the fully connected layer of the damage assessment model according to the dynamic weight matrix, generate a dynamic monitoring model and output the evaluation result of the deterioration risk of the rock mass structure.

[0013] In a preferred embodiment, in step S1, the rock mass acoustic wave propagation velocity data is collected by a cross-hole acoustic wave testing device, and the acoustic wave transmitters and receivers are arranged at intervals along the strike of the mining face, and the frequency range is adapted to the scale of the rock mass fissures;

[0014] The microseismic event energy spectrum data is collected by an underground microseismic monitoring network arranged in a honeycomb topology, and the geophone array covers the goaf boundary and meets the requirements of the spatial sampling density;

[0015] The borehole imaging data is collected by a borehole television imaging system, and the resolution and frame rate of the imaging device match the observation requirements of the rock mass fracture zone;

[0016] The rock mass acoustic wave propagation velocity data, the microseismic event energy spectrum data and the borehole imaging data are fused into a dynamic damage increment data set after time stamp alignment.

[0017] In a preferred embodiment, in step S2, the new damage features are extracted from the dynamic damage increment data set by a feature encoder including a convolutional layer and a fully connected layer, and the convolutional kernel size of the feature encoder is adapted to the spatial distribution characteristics of the rock mass damage;

[0018] Based on the mean and covariance matrix of each historical feature vector in the historical damage feature library, calculate the Mahalanobis distance of the new damage features as a similarity measure;

[0019] The dynamic distribution shift degree is generated by comparing the Mahalanobis distance after normalization with a preset similarity threshold, where the normalization coefficient is the reciprocal of the number of samples in the historical feature library;

[0020] When the new damage features include the sudden drop of the acoustic wave velocity and the synchronous enhancement mode of the high-frequency components of the microseismic energy spectrum, a time series correlation weighting factor is introduced in the calculation of the similarity measure.

[0021] In a preferred embodiment, the value of the time series correlation weighting factor is positively correlated with the mining advance rate.

[0022] In a preferred embodiment, when the dynamic distribution offset in step S3 exceeds the offset threshold, the mechanical behavior constraint verification of the newly added damage features based on the rock mass constitutive equation includes: verifying whether the stress state corresponding to the damage features is within the shear strength envelope of the rock mass through the Mohr-Coulomb criterion, and eliminating abnormal features beyond the strength envelope;

[0023] According to the law of conservation of energy in the damage evolution process, calculate the energy coupling coefficient of microseismic energy release and acoustic wave velocity change, and eliminate non-physical features with energy coupling coefficients deviating from the preset range;

[0024] The screening criterion for the effective damage feature subset is the feature set that simultaneously satisfies the stress state constraint and the energy conservation constraint;

[0025] When the mining stage enters the period of periodic weighting, the friction angle parameter of the stress state constraint is dynamically corrected according to the real-time in-situ stress monitoring data.

[0026] In a preferred embodiment, in step S4, performing damage path-dependent modeling on the effective damage feature subset includes: extracting the rock mass fracture propagation direction and rate parameters to construct a path similarity matrix according to the damage evolution pattern in the same geological structure area in the historical damage feature library;

[0027] Generate a spatial gradient direction field based on the spatio-temporal distribution characteristics of the effective damage feature subset, and the gradient direction of each node in the direction field is determined by the difference vector of the damage degrees of adjacent nodes;

[0028] The potential failure path heat map is generated through the convolution operation of the path similarity matrix and the spatial gradient direction field;

[0029] When the mining face reveals a fault structure, the weight coefficient of the path similarity matrix is corrected according to the cosine value of the angle between the fault strike and the current gradient direction.

[0030] In a preferred embodiment, the convolution kernel size of the convolution operation is dynamically adjusted according to the rock mass anisotropy parameter.

[0031] In a preferred embodiment, in step S5, using the knowledge distillation mechanism to perform cross-stage correlation analysis between the potential failure path heat map and the historical damage feature library includes: extracting the key attention areas of the temporal damage evolution pattern in the historical damage feature library through the teacher model, and extracting the spatio-temporal saliency features of the potential failure path heat map through the student model;

[0032] Match the attention weights of the teacher model with the saliency features of the student model in terms of spatial position, and the matching degree is calculated as the cosine similarity of the feature vectors at the corresponding positions;

[0033] The generation of the dynamic weight matrix includes the product of a time decay factor and a spatial correlation weight. The time decay factor decays exponentially according to the time interval between mining stages, and the spatial correlation weight is determined based on the linear combination of the matching degree and the thermal value;

[0034] When there is a fault structure in the mining area, the spatial correlation weight is corrected according to the cosine value of the angle between the fault strike and the thermal gradient direction.

[0035] In a preferred embodiment, when step S6 re-weights the parameters of the fully connected layer of the damage assessment model according to the dynamic weight matrix, each element of the dynamic weight matrix is multiplied by the connection weight of the corresponding neuron in the fully connected layer, and the product result is used as the updated connection weight;

[0036] The updated connection weight is restricted in terms of the change amplitude through a gradient clipping operation, and the clipping threshold is set based on the statistical fluctuation range of the historical weight parameters;

[0037] The generation of the dynamic monitoring model requires verifying the convergence of the model after parameter update. The convergence criterion is that the decrease amplitude of the loss function value of the model on the validation set is lower than the set ratio for consecutive multiple iterations;

[0038] The output of the rock mass structure deterioration risk assessment result includes the risk level division and spatial position marking of the mining area. The risk level is mapped to a preset risk interval through the weighted summation result of the dynamic weight matrix, and the spatial position marking is strictly aligned with the spatial grid coordinates of the mining face.

[0039] On the other hand, the present invention provides a mining-induced rock mass structure deterioration monitoring system based on deep learning, including the following modules:

[0040] A dynamic acquisition module, used to collect multi-source heterogeneous monitoring data under mining disturbances in real time and construct a dynamic damage increment data set;

[0041] A feature extraction module, used to extract the new damage features of the current mining stage based on the dynamic damage increment data set, and generate a dynamic distribution deviation degree based on the similarity measurement between the new damage features and the historical damage feature library;

[0042] A mechanical verification module, used to perform mechanical behavior constraint verification on the new damage features when the dynamic distribution deviation degree exceeds the deviation degree threshold, and screen out an effective damage feature subset that conforms to the mechanical laws of rock mass failure;

[0043] A path modeling module, used to perform damage path dependence modeling on the effective damage feature subset, match the damage evolution pattern in the same geological structure area in the historical damage feature library, and generate a potential failure path thermal map based on the spatial gradient;

[0044] A knowledge distillation module, which is used to perform cross-stage correlation analysis on the potential damage path heat map and the historical damage feature library by using a knowledge distillation mechanism, and generate a dynamic weight matrix reflecting the mapping relationship between the mining time sequence and the damage evolution;

[0045] A dynamic update module, which is used to re-weight the parameters of the fully connected layer of the damage assessment model according to the dynamic weight matrix, generate a dynamic monitoring model and output the evaluation result of the rock mass structure deterioration risk.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. Through a deep collaboration mechanism that dynamically fuses multi-source monitoring data and physical and mechanical laws, a closed-loop feedback system for rock mass damage evolution is constructed. Based on real-time mining disturbance data, a dynamic damage increment data set is constructed, breaking through the dependence of the traditional static training set on the historical data distribution. By incremental learning, the spatio-temporal evolution characteristics of newly generated fractures are continuously captured, effectively solving the problem of model inaccuracy caused by the irreversibility of rock mass damage. A mechanical behavior constraint verification mechanism is introduced to screen constitutive equations for newly added damage features, blocking the interference of non-physical damage modes, and ensuring that the model update always conforms to the mechanical mechanism of rock mass fracture, significantly improving the physical credibility of damage assessment. By coupling geological structure matching and spatial gradient analysis, historical damage patterns are transformed into spatio-temporal constraints of dynamic evolution paths, realizing the organic integration of empirical data and real-time monitoring, and avoiding the monitoring blind spots of traditional methods in new geological structure areas.

[0048] 2. By establishing cross-stage spatio-temporal associations through a knowledge distillation mechanism, mining time sequence information is encoded into the dynamic weight matrix, endowing the model with the dynamic adaptation ability to the rock mass deterioration process. The parameter re-weighting process combines gradient clipping and dynamic normalization strategies, achieving precise regulation of the parameter space while maintaining the model stability, forming a dual constraint of data-driven and physical laws, enabling the monitoring model to autonomously adapt to the feature distribution shift caused by mining disturbances and correcting the evaluation deviation in real time. Through the deep coupling of physical mechanisms and data features, an interpretable model of the rock mass damage evolution process is realized in a complex mining environment, providing decision support with both real-time and reliability for mine safety monitoring, and effectively solving the problems of early warning lag and misjudgment caused by model rigidity in traditional methods. Description of the Drawings

[0049] Figure 1 It is a flow chart of the method for monitoring the deterioration of the mined rock mass structure based on deep learning of the present invention;

[0050] Figure 2 It is a structural schematic diagram of the system for monitoring the deterioration of the mined rock mass structure based on deep learning of the present invention. Detailed Embodiments

[0051] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0052] Embodiment 1: Figure 1 A method for monitoring the deterioration of the mined rock mass structure based on deep learning according to the present invention is given, which includes the following steps:

[0053] S1. Real-time collect multi-source heterogeneous monitoring data under mining disturbances and construct a dynamic damage increment data set;

[0054] S2. Extract the newly added damage features in the current mining stage based on the dynamic damage increment data set, and generate a dynamic distribution deviation degree based on the similarity measurement between the newly added damage features and the historical damage feature library;

[0055] S3. When the dynamic distribution deviation degree exceeds the deviation degree threshold, perform mechanical behavior constraint verification on the newly added damage features, and screen out an effective damage feature subset that conforms to the mechanical laws of rock mass failure;

[0056] S4. Perform damage path dependence modeling on the effective damage feature subset, match the damage evolution patterns in the historical damage feature library in the same geological structure area, and generate a potential failure path heat map based on the spatial gradient;

[0057] S5. Adopt a knowledge distillation mechanism to perform cross-stage correlation analysis on the potential failure path heat map and the historical damage feature library, and generate a dynamic weight matrix reflecting the mapping relationship between the mining time sequence and the damage evolution;

[0058] S6. Re-weight the parameters of the fully connected layer of the damage assessment model according to the dynamic weight matrix, generate a dynamic monitoring model and output the risk assessment result of the rock mass structure deterioration.

[0059] S1. Real-time collect multi-source heterogeneous monitoring data under mining disturbances and construct a dynamic damage increment data set, including:

[0060] The multi-source heterogeneous monitoring data includes rock mass acoustic wave propagation velocity data, microseismic event energy spectrum data, and borehole imaging data.

[0061] When the cross-hole acoustic wave testing device collects the acoustic wave propagation velocity data of the rock mass, the layout spacing between the acoustic wave transmitter and the receiver is dynamically adjusted according to the integrity of the rock mass. The integrity coefficient of the rock mass is calculated through the Rock Quality Designation (RQD) value of the borehole core. The RQD value is the percentage of the cumulative length of the intact core with a length greater than 10 cm in the total length of the borehole. For example, when the RQD value is greater than 75%, the spacing between the transmitter and the receiver is set to 50 m. When the RQD value is less than 75%, the spacing is shortened to 30 m to adapt to the acoustic wave attenuation characteristics of rock masses with different integrities. The selection of the acoustic wave frequency range is based on the target scale of the crack to be detected, calculated by v / (2d), where v is the longitudinal wave velocity of the rock mass (in m / s) and d is the average opening width of the crack (in m). For example, the lower frequency limit of 10 kHz corresponds to detecting cracks larger than 5 mm, and the upper limit of 50 kHz corresponds to detecting cracks at the 1 mm level.

[0062] The honeycomb topology of the downhole microseismic monitoring network is arranged based on the goaf geometry and stress distribution characteristics. A single honeycomb unit is a regular hexagon, and its side length is dynamically adjusted according to the mining stage: for example, the side length is set to 30 m in the initial mining stage to cover a large range of low-stress areas, and shortened to 20 m in the periodic weighting stage to focus on high-stress concentration areas. The range of the three-component acceleration sensor of the geophone node is determined according to the energy of the largest historical microseismic event. For example, the upper limit of the range is 1.5 times the peak acceleration of the largest historical microseismic event. The microseismic event trigger threshold is set through statistical analysis of the background noise. For example, it takes 3 times the standard deviation of the root mean square value of the background noise acceleration in 24 consecutive hours. The calculation method is to calculate the root mean square value of the noise signal in each 5-minute time window and take 3 times the maximum value within 24 hours as the threshold.

[0063] The resolution of the imaging device and the brightness of the light source of the borehole television imaging system are set according to the borehole diameter and the light transmittance of the rock mass. For example, when the borehole diameter is less than 75 mm, a resolution of 1920×1080 pixels is used, and when the borehole diameter is greater than 75 mm, it is increased to 2560×1440 pixels. The brightness of the light source is adaptively adjusted based on the ambient light intensity in the hole. The adjustment algorithm includes setting the initial brightness to 2000 lumens, increasing the brightness by 500 lumens for each 1 m decrease in the borehole depth. If the dust concentration sensor in the hole detects that the light transmittance is lower than 70%, an additional 1000 lumens of brightness is added. The calibration scale verification in the crack width calculation is achieved through the reflective marks preset in the hole, and the interval error of the marks is controlled within ±1 mm. For example, a reflective mark is set every 50 cm.

[0064] During the timestamp alignment process, the system clock of the microseismic monitoring network serves as the master clock. The acoustic data acquisition device synchronizes with the master clock through the NTP protocol. The clock deviation compensation adopts a linear regression algorithm. For example, the clock drift rate is calculated every 5 minutes and dynamically corrected. The timestamp of the borehole imaging data is calculated by back-calculating the probe lifting speed and hole depth. The lifting speed error is feedback-controlled in a closed loop through an encoder. For example, the speed fluctuation range does not exceed ±5% of the set value. The aligned multi-source data is sliced and fused according to a time window. The length of the time window is set according to the rock mass damage evolution rate. For example, when the damage rate is lower than 0.1 mm / h, the window length is 10 minutes, and when it is higher than 0.1 mm / h, the window length is shortened to 5 minutes.

[0065] In the data normalization process of the dynamic damage increment data set, the normalization reference value of the acoustic wave velocity is dynamically adjusted according to the lithology. For example, 5500 m / s is used as the reference for granite-like rock masses, and 4500 m / s is used for sandstone-like rock masses. The logarithmic transformation base of the microseismic energy spectrum is selected according to the background noise level of the monitoring area. For example, the natural logarithm is used when the noise level is lower than 0.01 m / s², and the logarithm with base 10 is used when it is higher than 0.01 m / s² to suppress noise interference. The normalization threshold of the borehole fracture width is determined by the statistical value of the historical maximum fracture width. For example, the historical maximum value is re-statistically calculated every 50 meters of mining distance advanced.

[0066] The abnormal data processing mechanism includes the abnormal determination rule of acoustic signals: if the deviation of the continuous 3 wave velocity measurement values of the same transmitter-receiver pair exceeds 15%, it is determined as a device failure. The abnormal detection of the microseismic signal channel adopts the adjacent node data correlation analysis method. For example, when the correlation coefficient is lower than 0.7, it is marked as an abnormal channel. The evaluation of the borehole image clarity is based on the calculation of the Laplacian gradient operator. For example, when the gradient value is less than the set threshold, the re-acquisition mechanism is triggered. The data interpolation method is selected according to the data type. For example, the spatial Kriging interpolation is used for the acoustic wave velocity, the time linear interpolation is used for the microseismic energy spectrum, and the interpolation processing of the borehole fracture data is prohibited.

[0067] S2. Extract the new damage features of the current mining stage based on the dynamic damage increment data set, and generate the dynamic distribution deviation degree based on the similarity measurement between the new damage features and the historical damage feature library, including:

[0068] The feature encoder includes a convolutional layer and a fully connected layer. The size of the convolutional kernel in the convolutional layer is set according to the spatial distribution characteristics of rock mass damage. For example, a 3×5 rectangular convolutional kernel is used for the lateral damage expansion along the strike of the mining face, and a 5×3 rectangular convolutional kernel is used for the longitudinal damage in the direction perpendicular to the rock strata. The stride of the convolutional kernel is set to 1×1 to ensure the resolution of the feature map. The padding mode of the convolutional layer is selected as "valid", and the convolution operation is only performed within the range allowed by the input data size. The number of neurons in the fully connected layer is consistent with the dimension of the historical damage feature library. For example, when the dimension of the historical feature library is 256, the number of output neurons in the fully connected layer is set to 256. The ReLU function is used as the activation function to retain the non-linear damage features, and a Dropout layer is added after the fully connected layer to prevent overfitting. The Dropout ratio is set to 0.2. The dynamic damage increment data set is subjected to zero-mean normalization before being input into the feature encoder. The normalization parameters are the mean and standard deviation of all samples in the historical damage feature library. For example, the mean of the acoustic wave velocity feature is 4000m / s, and the standard deviation is 500m / s.

[0069] The mean and covariance matrix of each historical feature vector in the historical damage feature library are obtained through the offline training stage. The mean is calculated as the arithmetic mean of each dimension of the historical feature vector. For example, the average value of each dimension of 1000 historical feature samples is calculated separately. The covariance matrix is calculated as the mean of the outer product of the deviation of the historical feature vector from its mean. For example, the size of the covariance matrix of a feature vector with a dimension of 256 is 256×256. When calculating the Mahalanobis distance of the newly added damage feature, the inverse matrix of the covariance matrix is obtained by the singular value decomposition method. The minimum singular value threshold is set to 10 during the singular value decomposition process -6 When the singular value is less than this threshold, it is truncated to avoid numerical instability. The similarity measure is the reciprocal of the Mahalanobis distance, and the calculation formula is 1 / (1 + Mahalanobis distance), so as to map the measure value to the interval from 0 to 1. For example, when the Mahalanobis distance is 2, the similarity measure value is 0.333.

[0070] During the singular value decomposition process of the covariance matrix, let the covariance matrix be Its singular value decomposition is:

[0071] ;

[0072] Among them, represents the covariance matrix of the historical damage feature library, with a dimension of is the feature dimension; and are orthogonal matrices of ( is the identity matrix); is a diagonal matrix, ​denotes the singular value; is the singular value truncation threshold.

[0073] The newly added damage feature vector has the following Mahalanobis distance formula:

[0074] ;

[0075] where, denotes the Mahalanobis distance of the newly added damage feature vector ; denotes the current newly added damage feature vector, with a dimension of ; denotes the mean vector of the historical feature library, with a dimension of ; denotes the inverse matrix of the covariance matrix, calculated after being corrected by singular value decomposition.

[0076] In the normalization process of the dynamic distribution offset degree, the normalization coefficient is taken as the reciprocal of the number of samples in the historical damage feature library. For example, when the number of samples is 1000, the normalization coefficient is 0.001; the preset similarity threshold is determined by the 95th percentile of the historical data distribution. The specific method is to sort the historical similarity measurement values in ascending order and take the value corresponding to the 95th position as the threshold. For example, when the historical similarity measurement sequence is [0.1, 0.2,..., 0.9], the 95th percentile is 0.895; when generating the dynamic distribution offset degree, if the normalized Mahalanobis distance exceeds the similarity threshold, it is determined that there is a significant distribution offset, and the offset degree value is set to 1, otherwise it is the ratio of the normalized Mahalanobis distance to the threshold. For example, when the normalized Mahalanobis distance is 0.8 and the threshold is 0.7, the offset degree value is 0.8 / 0.7 ≈ 1.14, but after truncation processing, the final value is 1.0.

[0077] When the sudden drop in the acoustic wave velocity in the newly added damage feature exceeds 20% of the historical maximum drop and the energy ratio of the microseismic energy spectrum in the frequency band above 100 Hz increases by more than 30%, it is determined as the synchronous enhancement mode; the time series correlation weighting factor is calculated by linear interpolation according to the advancing rate of the mining face. For example, when the advancing rate is 3 meters per day, the weighting factor is 1.0, and when the rate increases to 5 meters per day, the weighting factor is adjusted to 1.5; the calculation formula of the weighted similarity measurement is the original similarity measurement value multiplied by the weighting factor. When the weighting factor is greater than 1, the result needs to be truncated, and the maximum value does not exceed 1.0. For example, when the weighting factor is 1.5 and the original similarity measurement value is 0.8, the calculation result is 1.2, and after truncation, it takes 1.0.

[0078] The training process of the feature encoder is completed in the offline stage. The training data is all samples in the historical damage feature library. The mean squared error function is used as the loss function, and the Adam algorithm is selected as the optimizer. The initial learning rate is set to 0.001 and decays to 90% of the original value every 50 training epochs. The training termination condition is that the loss decrease amplitude of the validation set for 10 consecutive epochs is less than 1%. The final model parameters are saved as the initial weights of the feature encoder. The training hardware environment needs to support CUDA parallel computing. For example, using NVIDIA Tesla V100 GPU to accelerate the training process, and the software dependencies include TensorFlow version 2.4 and above.

[0079] The exception handling mechanism includes the detection and repair of the singularity of the covariance matrix. When the matrix condition number is greater than 10 6 it is determined to be near singular, and at this time, a unit matrix of order 10 -5 is added by diagonal loading to enhance stability. If negative values appear during the calculation of the Mahalanobis distance, the data integrity check is immediately triggered, and the feature vector of the previous time series is rolled back for recalculation. A moving average filter with a window length of 5 time steps is performed before the output of the dynamic distribution deviation degree to suppress instantaneous noise interference. If the input data dimension does not match during the inference stage of the feature encoder, dimension filling or truncation is automatically triggered, and the filling value is the global mean of the corresponding feature.

[0080] The boundary condition handling includes the coping strategies for extreme input data. For example, when the sudden drop amplitude of the acoustic wave velocity exceeds 50%, it is determined as a sensor failure and switched to the backup data channel. When the microseismic energy spectrum appears saturated in the full frequency band, dynamic gain adjustment is started, and the sensor range is temporarily expanded to 2 times. The covariance matrix update period is set to recalculate every 100 new feature samples to ensure the timeliness of the matrix. The upper limit of the capacity of the historical damage feature library is set to 10,000 samples, and the old samples are replaced according to the first-in, first-out principle after reaching the limit.

[0081] S3. When the dynamic distribution deviation degree exceeds the deviation degree threshold, the mechanical behavior constraint verification of the newly added damage features is carried out to screen out the effective damage feature subset that conforms to the mechanical law of rock mass failure, including:

[0082] When the dynamic distribution offset exceeds the offset threshold, the mechanical behavior constraint verification of the new damage characteristics is carried out based on the rock mass constitutive equation. First, it is verified whether the stress state corresponding to the damage characteristics is within the shear strength envelope of the rock mass through the Mohr-Coulomb criterion; the cohesion parameter and friction angle parameter of the Mohr-Coulomb criterion are dynamically set according to the rock mass lithology category, and the lithology category is determined by the mineral composition analysis results of borehole television imaging. The mineral composition analysis uses X-ray diffraction spectroscopy to identify the main mineral types and contents; the cohesion parameter is obtained through the uniaxial compression test of laboratory rock cores, and the test loading rate is 0.5 mm / min until the rock sample is damaged, and the peak stress is recorded as the cohesion value; the friction angle parameter is determined through the direct shear test. The normal stress of the direct shear test is applied in grades, and the shear displacement and shear stress are recorded under each normal stress, and the shear stress-displacement curve is plotted. The angle corresponding to the peak shear stress of the curve is taken as the friction angle; the shear strength envelope is calculated through the relational formula of normal stress and shear stress. The normal stress is the arithmetic mean of the maximum principal stress and the minimum principal stress in the damage characteristics, and the shear stress is half of the difference between the maximum principal stress and the minimum principal stress; if the calculated shear stress exceeds the theoretical value of the envelope corresponding to the current normal stress during verification, it is determined as a characteristic that violates the rock mass shear strength law and is excluded. At the same time, an abnormal characteristic report is generated, which includes abnormal stress values, verification time, and associated sensor numbers.

[0083] The determination of the dynamic distribution offset threshold needs to combine the statistical analysis of historical mining data. The specific implementation includes: collecting historical dynamic distribution offset data for at least three months in the initial stage, and the data covers different mining intensities (such as daily advance rate of 1-5 meters) and geological conditions (such as fault zones, intact rock strata); performing the Kolmogorov-Smirnov test on the data distribution. If the test P value is greater than 0.05, it is determined to be a normal distribution, and the threshold is set according to the mean plus three standard deviations, otherwise it is set according to the 95% quantile; for example, the historical offset data of a certain mining area is tested as non-normal distribution, and the 95% quantile after sorting is 0.85, then the initial threshold is set to 0.85; the threshold is updated once a month, and when updating, the offset data of the latest 30 days is included to recalculate the quantile or mean standard deviation; when the geological conditions of the mining area change significantly (such as the discovery of a large fault), the threshold recalculation is immediately triggered and manually reviewed and confirmed.

[0084] According to the law of conservation of energy during the damage evolution process, calculate the energy coupling coefficient between the microseismic energy release and the change in acoustic wave velocity. The calculation logic of the energy coupling coefficient is the ratio of the energy released by microseismic events to the energy consumed by the change in acoustic wave velocity. The microseismic energy is obtained by integrating the area of the microseismic event energy spectrum within a time window. The integration interval is from 10 seconds before the event trigger time to 20 seconds after, and the frequency resolution is 1 Hz. The energy consumed by the change in acoustic wave velocity is calculated by multiplying the wave velocity change rate by the rock mass density. The wave velocity change rate is obtained by dividing the wave velocity difference between adjacent time windows by the time interval, and the rock mass density is determined by the laboratory measurement value of the borehole core. The preset interval is determined by the data statistics of the historical normal mining stage. The statistical method is to calculate the mean of the historical energy coupling coefficient plus or minus twice the standard deviation. The historical data covers at least three complete mining cycles. The dimension of the energy coupling coefficient is a dimensionless ratio. During the calculation process, the unit of microseismic energy is converted to joules, the unit of acoustic wave velocity is unified to meters per second, and the unit of rock mass density is kilograms per cubic meter. All unit conversion coefficients are calibrated uniformly before data processing.

[0085] The screening criteria for the effective damage feature subset are the feature set that simultaneously satisfies the stress state constraint and the energy conservation constraint. The implementation method is to sequentially perform stress state verification and energy coupling verification on each newly added damage feature, and only retain the features that pass both verifications. The result of the stress state verification is recorded by an independent status flag bit. The flag bit being 1 indicates passing the verification, and 0 indicates not passing. The energy coupling verification is recorded by another independent flag bit. The generation rule of the flag bit is to set it to 1 when the energy coupling coefficient is within the preset interval, otherwise set it to 0. Finally, the comprehensive verification flag bit is generated through a logical AND operation. The comprehensive flag bit is set to 1 only when both flag bits are 1. The feature set with the comprehensive flag bit being 1 is transmitted to the subsequent processing link, and the features with the flag bit being 0 are stored in the abnormal feature library. The abnormal feature library is indexed and stored by timestamp and spatial location, supporting quick positioning during manual review.

[0086] When the mining stage enters the periodic weighting period, the friction angle parameter of the stress state constraint is dynamically corrected according to the real-time in-situ stress monitoring data. The dynamic correction method is as follows: The real-time in-situ stress data is collected by borehole stress gauges. The borehole stress gauges are arranged in a fan shape on the roof of the goaf. The vertical in-situ stress is the reading of the vertical component of the stress gauge, and the horizontal in-situ stress is the reading of the maximum horizontal principal stress component. The in-situ stress ratio is calculated by the ratio of the horizontal in-situ stress to the vertical in-situ stress.

[0087] Among them, the friction angle correction formula is:

[0088] ;

[0089] Among them, represents the corrected friction angle (unit: degree); represents the initial friction angle (unit: degree), which is determined by laboratory direct shear tests; denotes the correction coefficient (unit: degrees / unit in-situ stress ratio), which is determined by linear regression analysis of the in-situ stress ratio and the change in friction angle in rock mechanics tests, and the value range is ; denotes the current in-situ stress ratio, is the horizontal in-situ stress (unit: MPa), is the vertical in-situ stress (unit: MPa), which is measured in real-time by a borehole stress gauge; denotes the reference value of the in-situ stress ratio, which is calculated by the average value of historical in-situ stress ratio data during the mine development stage,

[0090] The calculation formula of

[0091] ;

[0092] wherein, is the number of historical data samples, is the number of the historical data sample.

[0093] The corrected friction angle is updated every 2 hours. If the in-situ stress data is missing, the correction value of the previous time period is used, and a data integrity alarm is triggered.

[0094] The exception handling mechanism includes the extreme stress handling rules in stress state verification. When the maximum principal stress exceeds 80% of the uniaxial compressive strength of the rock mass, the manual review process is automatically triggered and the automated screening is suspended; if the acoustic wave velocity change rate is zero in the calculation of the energy coupling coefficient, the calculation of the current time window is skipped and marked as data invalid, and the invalid data segment is marked as "requiring manual verification" in the database; when the in-situ stress data is missing during the dynamic correction of the friction angle, the sliding average method is used to interpolate and complete the data based on the data of the previous 6 hours, and the interpolation weights are assigned as [0.4, 0.3, 0.2, 0.1] according to the proximity of time; when all exception events are triggered, a timestamped log record is generated, and the log includes the exception type, associated feature index, handling measures and operator signature fields, and the log files are archived daily and encrypted for storage.

[0095] Boundary condition coverage includes the adaptive adjustment rule for the energy coupling coefficient interval. When the number of effective features in five consecutive time windows is lower than 50% of the historical mean, the interval range is expanded by a preset step size. The expansion step size is 10% of the original interval width, and the maximum expansion amplitude does not exceed 50% of the original interval. The value range of the normal stress in the stress state verification is limited to 0 to 80% of the uniaxial compressive strength of the rock mass. Features outside the range are directly discarded and recorded as "overlimit data". The friction angle correction value is forcibly restricted within the range allowed by the physical properties of the rock mass. If the calculated value exceeds the maximum friction angle corresponding to the rock mass type (for example, the upper limit of sandstone is 45 degrees, and the upper limit of granite is 55 degrees), it is truncated according to the maximum value. The truncation operation is recorded in the log and triggers a system notification.

[0096] S4. Perform damage path-dependent modeling on the effective damage feature subset, match the damage evolution pattern in the same geological structure area in the historical damage feature library, and generate a thermal map of potential failure paths based on the spatial gradient, including:

[0097] When performing damage path-dependent modeling on the effective damage feature subset, according to the damage evolution pattern in the same geological structure area in the historical damage feature library, extract the rock mass fracture propagation direction and rate parameters to construct a path similarity matrix. The matching method for the same geological structure area is as follows: Calculate the Euclidean distance between the geological structure parameters of the current mining area, including fault strike, dip angle, and rock layer thickness, and the geological structure parameters recorded in the historical feature library, and select the top five historical areas with the smallest Euclidean distance as the reference patterns. The fracture propagation direction parameter is determined by the mean azimuth angle of the fracture traces in the historical area. The azimuth angle of the fracture trace is measured as the angle between the straight line fitted to the fracture contour by borehole television imaging and the due north direction. The rate parameter is calculated by the change rate of the fracture length over time. The change rate is the ratio of the fracture length difference to the time interval, and the time interval is dynamically adjusted according to the mining stage. It is 24 hours in the initial mining stage and shortened to 6 hours in the periodic weighting stage. Each element of the path similarity matrix represents the similarity between the current damage feature and the historical pattern in terms of direction and rate. The direction similarity is calculated as the cosine value of the angle between the current fracture direction and the historical fracture direction, and the rate similarity is calculated as the ratio of the current rate to the historical rate and restricted to between 0.1 and 10. The final similarity is the product of the direction cosine value and the rate ratio.

[0098] Generate a spatial gradient direction field based on the spatio-temporal distribution characteristics of the effective damage feature subset. The gradient direction of each node in the spatial gradient direction field is determined by the difference vector of the damage degrees of adjacent nodes; the spatial grid division is set according to the advancing direction of the mining face, the grid cell size is 5 meters × 5 meters, and adjacent nodes are defined as the six nodes directly connected in space, namely up, down, left, right, front, and back; the difference vector is calculated as the difference vector between the damage degree of the current node and the damage degrees of adjacent nodes, the difference direction points from the high-damage area to the low-damage area, and the vector modulus is the absolute value of the damage degree difference; during the generation process of the gradient direction field, if the data of adjacent nodes is missing, the inverse distance weighted interpolation method is used to complete the data, and the interpolation weight is inversely proportional to the square of the distance between nodes. For example, when the distance between adjacent nodes is 2 meters, the weight is 0.25; the gradient direction field is updated every 30 minutes. When updating, the direction field of the previous time period is retained as the reference benchmark. If the direction deviation between the current direction field and the benchmark field exceeds 30 degrees, a direction anomaly alarm is triggered and manual verification is started.

[0099] The thermal map of potential failure paths is generated through the convolution operation of the path similarity matrix and the spatial gradient direction field. The size of the convolution kernel is dynamically adjusted according to the rock mass anisotropy parameter; the rock mass anisotropy parameter is obtained through the wave velocity anisotropy test of laboratory rock cores. During the test, acoustic waves are applied along the axial, tangential, and normal directions of the rock core and the wave velocities are recorded. The calculation formula for the anisotropy difference ratio is (maximum wave velocity - minimum wave velocity) / average wave velocity × 100%; the transverse size (along the mining direction) of the convolution kernel is set as a linear function of the anisotropy difference ratio. For example, when the difference ratio is 20%, the transverse size is 7×3 and the longitudinal size is 3×7.

[0100] The zero-padding mode is adopted for the convolution operation to keep the size of the output thermal map unchanged. The zero-padding convolution formula is:

[0101] ;

[0102] where, represents the original convolution result at position in the output thermal map; represents the weight parameter of the convolution kernel at position and input channel ; represents the value of the input data at position and channel ; represents the number of zero-padding layers. In the zero-padding mode, zeros are padded at the edges of the input data to keep the output size unchanged; represents the number of input channels; is the size of the convolution kernel, is the height of the convolution kernel, is the width of the convolution kernel.

[0103] The convolution result is normalized to the range of 0 to 1 through the Sigmoid function, and the normalization formula is:

[0104] ;

[0105] where is the original value of the convolution result; represents the normalized heat value, and the value range is [0, 1]; is the base of the natural logarithm (Euler's number), ; When it is judged as a high-risk area, a warning signal is triggered. The threshold of 0.8 is set according to the statistics of historical accident data.

[0106] The closer the value in the heat map is to 1, the higher the potential damage risk. When the heat value exceeds 0.8, it is marked as a high-risk area and a warning signal is triggered.

[0107] When the mining face reveals a fault structure, the weight coefficient of the path similarity matrix is corrected according to the cosine value of the angle between the fault strike and the current gradient direction; the fault strike is determined by the joint interpretation of borehole television imaging and ground penetrating radar data. The interpretation method is to extract the angle between the extension direction of the fault interface and the due north direction, and the interpretation accuracy is controlled within ±5 degrees; the cosine value of the angle is calculated as the dot product of the fault strike vector and the gradient direction vector divided by the product of the moduli of the two vectors, and the calculation result is limited between -1 and 1; the weight coefficient correction formula is the original weight multiplied by (1 + cosine value), and the corrected weight coefficient is limited to the range of 0 to 1 through the truncation function. For example, when the corrected calculation result is 1.2, 1.0 is taken, and when the calculation result is -0.3, 0 is taken; the trigger condition for weight coefficient update is that the working face advances to within 50 meters of the fault influence zone, and it is recalculated every 10 meters of advancement. If the weight coefficient fluctuates by more than 50% in two consecutive calculations, the automatic correction is suspended and a request for manual intervention is generated.

[0108] The anomaly handling mechanism includes the singular value handling of the path similarity matrix. When the matrix condition number is greater than 10 6 , a unit matrix of the order of 10 -5 is added for regularization; during the spatial gradient direction field interpolation process, if more than 50% of the adjacent node data is missing, it is determined as an invalid area and the heat map generation is suspended, and recalculated after the data is restored; when the convolution kernel size is dynamically adjusted, if the anisotropy parameter exceeds the laboratory test range (for example, the difference ratio > 30%), the preset maximum convolution kernel size of 11×5 is adopted, and the over-limit event is recorded in the log; after the heat map is generated, if the heat value of the same area in three consecutive time windows continuously exceeds 0.9, a red warning is triggered and the manual verification process is started, and the verification result is fed back to the historical damage feature library to update the damage evolution pattern.

[0109] Boundary condition coverage includes the similarity range limit of the path similarity matrix. When the rate ratio exceeds 10, it is forced to be truncated to 10 and marked as abnormal data; when the magnitude of the node damage degree difference vector in the spatial gradient direction field exceeds twice the historical maximum value, it is determined that the sensor is abnormal and the backup data source is enabled.

[0110] S5. Adopt the knowledge distillation mechanism to conduct cross-stage correlation analysis between the potential damage path heat map and the historical damage feature library, and generate a dynamic weight matrix reflecting the mapping relationship between the mining time sequence and damage evolution, including:

[0111] When conducting cross-stage correlation analysis between the potential damage path heat map and the historical damage feature library using the knowledge distillation mechanism, the teacher model is constructed based on the time sequence damage evolution pattern in the historical damage feature library. The teacher model extracts key attention regions through the self-attention mechanism; the query vector of the self-attention mechanism is the timestamp encoding of the current mining stage, and the timestamp encoding is generated using the sine position encoding method, and the encoding dimension is consistent with the timestamp dimension of the historical damage features; the key vector is the timestamp encoding of the historical damage features, and the value vector is the historical damage feature vector; the attention weight is calculated as the weight distribution after normalizing the dot product of the query vector and the key vector through the Softmax function, and the top 10% high-weight regions in the weight distribution are marked as key attention regions; the input of the student model is the potential damage path heat map, and spatio-temporal saliency features are extracted through a three-dimensional convolutional layer. The size of the three-dimensional convolutional kernel is dynamically adjusted according to the number of days between mining stages. For example, when the interval is 7 days, the kernel size is 3×3×3, and when the interval is 30 days, it is 5×5×5. The convolutional stride is fixed at 1×1×1 to retain spatio-temporal details.

[0112] When spatially matching the attention weights of the teacher model with the saliency features of the student model, the grid division for spatial position matching is consistent with the spatial resolution of the potential damage path heat map, and the grid cell size is 5 meters × 5 meters, which is aligned with the sensor layout grid of the mining face; the matching degree is calculated as the cosine similarity between the teacher model attention weight vector and the student model saliency feature vector within the corresponding grid cell. The cosine similarity calculation formula is the dot product of the two vectors divided by the product of the vector magnitudes; if there is a fault structure within the grid cell, the matching degree needs to be multiplied by the fault influence factor, and the fault influence factor is determined according to the cosine value of the angle between the fault dip and the thermal gradient direction. The specific rule is: when the cosine value of the angle is greater than 0.8, the factor is 1.5; when the cosine value is between 0.5 and 0.8, the factor is 1.2; when the cosine value is less than 0.5, the factor is 0.8. This rule is set based on the statistical results of the correlation between the fault slip tendency and the stress direction in the rock mechanics experiment.

[0113] The generation of the dynamic weight matrix includes the product of the time decay factor and the spatial correlation weight. The time decay factor is calculated according to the exponential decay law based on the time interval of the mining stage. Its base is the base e of the natural logarithm, and the exponential part is the product of the negative decay coefficient λ and the number of days Δt of the stage interval. The decay coefficient λ is determined according to the fitting curve of the damage evolution rate and the time decay effect in the historical mining data. Specifically, the corresponding relationship between the damage rate and the time decay is fitted by the least squares method. For example, when the damage rate is 0.1 mm per day, λ is taken as 0.01; when the damage rate increases to 0.5 mm per day, λ is adjusted to 0.05. The number of days Δt of the stage interval is the time difference between the current mining stage and the historical reference stage, and is statistically counted in natural days. The value range of λ is based on the regression analysis results of historical data to ensure the self-adaptability of the model to the mining process. The spatial correlation weight is determined by the linear combination of the matching degree and the thermal value. The linear combination coefficient is obtained by multiple linear regression analysis of historical data. The input variables of the regression model are the matching degree and the thermal value, and the output variable is the actually observed damage propagation probability. The regression results show that the weight ratio of the matching degree is 60% and the weight ratio of the thermal value is 40%. Each element value of the dynamic weight matrix is normalized to the interval of 0 to 1 through the Sigmoid function. When it exceeds 1, it is truncated to 1, and when it is lower than 0, it is set to 0, and the truncation event is recorded in the exception log.

[0114] When there is a fault structure in the mining area, the correction of the spatial correlation weight is based on the cosine value of the angle between the fault strike and the thermal gradient direction. The fault strike is determined by the joint interpretation of the joint statistics of the borehole core and the ground penetrating radar scan data. The interpretation method is to extract the angle between the extension direction of the fault interface and the due north direction, and the interpretation error is controlled within ±3 degrees by taking the average of three measurements. The cosine value is calculated as the dot product of the unit vector of the fault strike and the unit vector of the thermal gradient direction, and the calculation result is limited between -1 and 1. The corrected spatial correlation weight is the original weight multiplied by (1 + cosine value). For example, when the original weight is 0.6 and the cosine value is 0.5, the corrected weight is 0.6×1.5 = 0.9. If the corrected weight exceeds 1, it is set to 1, and if it is lower than 0, it is set to 0. The correction operation triggers a log record, and the log includes the fault number, correction time, operator identification, and the weight values before and after correction.

[0115] The exception handling mechanism includes the alignment check of the feature dimensions between the teacher model and the student model. If the dimensions do not match, they are automatically adjusted to be consistent through zero-padding or truncation operations. The zero-padding value is set to 0, and the first N dimensions of the features are retained for the truncated part (N is the output dimension of the student model); during the generation of the dynamic weight matrix, if it is detected that more than 50% of the grid cell weights are 0, it is determined that the data is abnormal and the weight matrix of the previous time step is rolled back, and at the same time, the data quality inspection process is triggered; when calculating the fault influence factor, if the interpreted data is missing, the correction mechanism is paused and the uncorrected weights are used for continued operation, and the missing data window is marked as "to be supplemented"; the hardware dependencies include NVIDIA GPUs (such as Tesla V100) to accelerate matrix operations, and CUDA 11.3 and above versions and the corresponding drivers need to be configured. The software environment needs to install the PyTorch 1.9 framework and load the pre-trained self-attention model parameters, and the parameter file format is the.pt format supported by the official PyTorch.

[0116] The boundary condition coverage includes that the maximum number of days for the time decay factor is limited to 365 days. When it exceeds, the decay coefficient λ is forced to be reset to the default value of 0.02 to prevent numerical overflow; the linear combination coefficient of the spatial correlation weights is recalibrated once a month, and the calibration data source is the valid feature samples in the most recent three months. When the number of samples is insufficient, the coefficient of the previous month is used.

[0117] S6. Re-weight the parameters of the fully connected layer of the damage assessment model according to the dynamic weight matrix to generate a dynamic monitoring model and output the assessment result of the deterioration risk of the rock mass structure, including:

[0118] When re-weighting the parameters of the fully connected layer of the damage assessment model according to the dynamic weight matrix, each element of the dynamic weight matrix is multiplied element by element with the connection weight of the corresponding neuron in the fully connected layer, and the product result is used as the updated connection weight; the dimension of the dynamic weight matrix is consistent with the number of input neurons of the fully connected layer. For example, when the input dimension of the fully connected layer is 256, the size of the dynamic weight matrix is 256×256 to ensure that the weights of each neuron connection are dynamically adjusted; the parameter re-weighting operation is performed before the model backpropagation to ensure that the latest weight parameters are loaded during the forward calculation; the updated connection weights are restricted in terms of the change range through gradient clipping. The clipping threshold is set based on the statistical fluctuation range of the historical weight parameters. The specific method is to calculate the average fluctuation amplitude of the historical weights during the training stage, and set the clipping threshold to a fixed multiple of the average fluctuation amplitude. For example, when the historical average fluctuation amplitude is 0.2, the clipping threshold is 0.3; the gradient clipping adopts the global norm clipping strategy to limit the weight update step size not to exceed the threshold and prevent gradient explosion or disappearance.

[0119] The generation of the dynamic monitoring model requires verifying the convergence of the model after parameter update. The convergence criterion is that the decline amplitude of the loss function value of the model on the validation set is lower than the set ratio for consecutive multiple iterations. The validation set is composed of data in the historical damage feature library that has not participated in training. The data is divided in a ratio where the training set and the validation set are split according to a preset ratio, such as 8:2. The calculation of the decline amplitude of the loss function value is the relative change rate of the loss values of two adjacent iterations. The set ratio is dynamically adjusted according to the model complexity. For example, when the number of neurons in the fully connected layer exceeds 1000, the set ratio is 0.5%, and when it is less than 1000, it is 1%. If the decline amplitude of the loss value for consecutive multiple iterations is lower than the set ratio, it is determined that the model converges and the parameter update is locked. When the convergence standard is not met, it rolls back to the previous weight parameter and triggers the learning rate adjustment mechanism. The adjustment amplitude of the learning rate is dynamically set according to the change trend of the loss value.

[0120] The output of the rock mass structure deterioration risk assessment result includes the risk level division and spatial position annotation of the mining area. The risk level is mapped to a preset risk interval through the weighted sum result of the dynamic weight matrix. The weighted sum calculation is the sum of the product of the weight value corresponding to each spatial position in the dynamic weight matrix and its associated damage feature value. The sum result is normalized to a preset interval through linear transformation, such as 0 to 100. The preset rule for the risk interval is: the low-risk interval corresponds to the normalized value from 0 to 30, the medium-risk interval corresponds to 30 to 70, and the high-risk interval corresponds to 70 to 100. The spatial position annotation is strictly aligned with the spatial grid coordinates of the mining face. The grid coordinates are generated based on the geographical registration information of the borehole imaging data. The center point coordinates of each grid unit are calibrated by a high-precision GPS positioning device, and the positioning error is controlled within the allowable range, such as ±0.1 meters. The risk annotation result is superimposed on the 3D mine model to support interactive visual query.

[0121] The exception handling mechanism includes an automatic rollback strategy when parameter update fails. If the model convergence verification fails and the learning rate adjustment exceeds the preset number of times, it automatically rolls back to the initial weight and generates a request for manual intervention. If it is detected that more than the preset ratio of the weight change amount is truncated during the gradient clipping process, it is determined as a gradient anomaly event, pauses the parameter update, and restarts the model initialization process. Before the output of the risk assessment result, a spatial consistency check is required. If the risk levels of adjacent grid units span multiple intervals, it triggers spatial smoothing filtering processing. The size of the filtering window is dynamically adjusted according to the grid resolution. For example, a 3×3 filtering window is used for a 5-meter resolution grid. The hardware dependencies include a graphics processor that supports parallel computing, such as the NVIDIA Tesla series GPU. The software dependencies include a deep learning framework and a geographic information processing library, such as TensorFlow 2.6 and GeoPandas 0.10.

[0122] Boundary condition coverage includes numerical range constraints on the dynamic weight matrix. The updated weight parameters are forced to be restricted between -1 and 1, and values outside this range are compressed to this interval through a non-linear function. The maximum number of iterations for model convergence verification is set to a reasonable upper limit. If convergence is not achieved after exceeding the upper limit, the process is terminated and marked as model failure.

[0123] In this embodiment, a closed-loop feedback system for rock mass damage evolution is constructed through the dynamic fusion of multi-source heterogeneous monitoring data and the deep embedding of physical and mechanical constraints. In the data acquisition stage, a spatial topology optimization strategy for cross-hole acoustic wave and microseismic monitoring is adopted to overcome the defect of insufficient response of traditional sensor layout to the anisotropy of geological structures. A rock mass constitutive equation verification mechanism is introduced to block the interference of noise data on model updates by screening damage characteristics that conform to mechanical laws in real time. The damage path dependence modeling creatively couples historical geological structure matching and spatial gradient analysis, transforming static empirical data into spatio-temporal constraint conditions for dynamic evolution path prediction. The knowledge distillation mechanism integrates temporal decay and spatial correlation weights in cross-stage correlation analysis, breaking through the static allocation mode of traditional weight matrices. The parameter reweighting process realizes the balance of model parameter update stability and self-adaptability through the synergistic effect of gradient clipping and dynamic normalization. Each technical link forms an organic whole: data acquisition provides physically real inputs for feature extraction, mechanical verification ensures the credibility of the evolution mode, path modeling endows geological significance to spatio-temporal prediction, knowledge distillation establishes a cross-stage knowledge transfer channel, and dynamic update ensures the continuous adaptation of the model to the mining process. The multi-dimensional constraint coupling and dynamic feedback mechanism of this embodiment realizes the full-process deep integration of data-driven models with physical laws and geological characteristics in the field of mine rock mass monitoring.

[0124] Embodiment 2: Figure 2 The structural schematic diagram of the mining-induced rock mass structure degradation monitoring system based on deep learning according to the present invention is given. The mining-induced rock mass structure degradation monitoring system based on deep learning includes the following modules:

[0125] The dynamic acquisition module is used to collect multi-source heterogeneous monitoring data under mining disturbances in real time and construct a dynamic damage increment data set;

[0126] The feature extraction module is used to extract new damage features in the current mining stage based on the dynamic damage increment data set, and generate a dynamic distribution offset degree based on the similarity measurement between the new damage features and the historical damage feature library;

[0127] The mechanical verification module is used to perform mechanical behavior constraint verification on the new damage features when the dynamic distribution offset degree exceeds the offset degree threshold, and screen out an effective damage feature subset that conforms to the mechanical laws of rock mass failure;

[0128] A path modeling module for performing damage path - dependent modeling on the subset of effective damage features, matching the damage evolution patterns in the historical damage feature library for the same geological structure area, and generating a heat map of potential failure paths based on spatial gradients;

[0129] A knowledge distillation module for performing cross - stage correlation analysis between the heat map of potential failure paths and the historical damage feature library using a knowledge distillation mechanism, and generating a dynamic weight matrix reflecting the mapping relationship between mining time sequence and damage evolution;

[0130] A dynamic update module for re - weighting the parameters of the fully - connected layer of the damage assessment model according to the dynamic weight matrix, generating a dynamic monitoring model and outputting the assessment result of the deterioration risk of the rock mass structure.

[0131] The calculations involved in the embodiments are dimensionless numerical calculations. The preset parameters and threshold selections in the calculations are set by those skilled in the art according to the actual situation.

[0132] The above - mentioned embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented using software, the above - mentioned embodiments can be implemented in whole or in part in the form of a computer program product.

[0133] Those of ordinary skill in the art can realize that the modules and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or by a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application of the technical solution and the invention constraints. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0134] In addition, in each embodiment of the present application, the functional modules can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module.

[0135] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is only a logical function division. In actual implementation, there can be other division methods. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of devices or modules can be in electrical, mechanical, or other forms.

[0136] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claimed rights.

[0137] Finally, the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A monitoring method for the deterioration of the structure of mined rock masses based on deep learning, characterized in that, It includes the following steps: S1. Collect multi-source heterogeneous monitoring data under mining disturbances in real time and construct a dynamic damage increment data set; S2. Extract the newly added damage features in the current mining stage based on the dynamic damage increment data set, and generate a dynamic distribution deviation degree based on the similarity measurement between the newly added damage features and the historical damage feature library; S3. When the dynamic distribution deviation degree exceeds the deviation threshold, conduct a mechanical behavior constraint verification on the newly added damage features, and screen out an effective damage feature subset that conforms to the mechanical laws of rock mass failure; S4. Perform damage path dependence modeling on the effective damage feature subset, match the damage evolution pattern in the historical damage feature library in the same geological structure area, and generate a potential failure path heat map based on the spatial gradient; S5. Use a knowledge distillation mechanism to conduct cross-stage correlation analysis between the potential failure path heat map and the historical damage feature library, and generate a dynamic weight matrix reflecting the mapping relationship between mining time series and damage evolution; S6. Re-weight the parameters of the fully connected layer of the damage assessment model according to the dynamic weight matrix, generate a dynamic monitoring model and output the rock mass structure deterioration risk assessment result.

2. The method for monitoring the deterioration of the mined rock mass structure based on deep learning according to claim 1, wherein, In step S1, the rock mass acoustic wave propagation velocity data is collected through a cross-hole acoustic wave testing device, and the acoustic wave transmitter and receiver are arranged at intervals along the strike of the mining face, and the frequency range is adapted to the rock mass fracture scale; The microseismic event energy spectrum data is collected through an underground microseismic monitoring network arranged in a honeycomb topology, and the geophone array covers the goaf boundary and meets the airspace sampling density requirements; The borehole imaging data is collected through a borehole television imaging system, and the resolution and frame rate of the imaging device match the observation requirements of the rock mass fracture zone; The rock mass acoustic wave propagation velocity data, microseismic event energy spectrum data and borehole imaging data are fused into a dynamic damage increment data set after time stamp alignment.

3. The method for monitoring the deterioration of the mined rock mass structure based on deep learning according to claim 2, characterized in that, In step S2, the newly added damage features are extracted from the dynamic damage increment data set through a feature encoder including a convolutional layer and a fully connected layer, and the convolutional kernel size of the feature encoder is adapted to the spatial distribution characteristics of rock mass damage; Based on the mean and covariance matrix of each historical feature vector in the historical damage feature library, calculate the Mahalanobis distance of the newly added damage features as the similarity measurement; The dynamic distribution deviation degree is generated by comparing the normalized Mahalanobis distance with a preset similarity threshold, where the normalization coefficient is the reciprocal of the number of samples in the historical feature library; When the newly added damage features include the mode of sudden drop in acoustic wave velocity and synchronous enhancement of the high-frequency component of the microseismic energy spectrum, a time series correlation weighting factor is introduced in the similarity measurement calculation.

4. The method for monitoring the deterioration of the mined rock mass structure based on deep learning according to claim 3, wherein The value of the time series correlation weighting factor is positively correlated with the mining advancement rate.

5. The method for monitoring the deterioration of the mined rock mass structure based on deep learning according to claim 4, characterized in that, In step S3, when the dynamic distribution deviation degree exceeds the deviation threshold, the mechanical behavior constraint verification of the newly added damage features based on the rock mass constitutive equation includes: verifying whether the stress state corresponding to the damage features is within the shear strength envelope of the rock mass through the Mohr-Coulomb criterion, and eliminating abnormal features beyond the strength envelope; According to the law of conservation of energy in the damage evolution process, calculate the energy coupling coefficient between microseismic energy release and acoustic wave velocity change, and eliminate non-physical features with energy coupling coefficients deviating from the preset interval; The screening criteria for the effective damage feature subset are the feature set that simultaneously satisfies the stress state constraint and the energy conservation constraint; When the mining stage enters the periodic weighting period, the friction angle parameter of the stress state constraint is dynamically corrected according to the real-time in-situ stress monitoring data.

6. The method for monitoring the deterioration of the mined rock mass structure based on deep learning according to claim 5, characterized in that, Performing damage path-dependent modeling on the effective damage feature subset in step S4 includes: extracting the rock mass fracture propagation direction and rate parameters according to the damage evolution pattern in the same geological structure area in the historical damage feature library to construct a path similarity matrix; Generating a spatial gradient direction field based on the spatio-temporal distribution characteristics of the effective damage feature subset, and the gradient direction of each node in the direction field is determined by the difference vector of the damage degrees of adjacent nodes; The potential failure path heat map is generated through the convolution operation of the path similarity matrix and the spatial gradient direction field; When the mining face reveals a fault structure, the weight coefficient of the path similarity matrix is corrected according to the cosine value of the angle between the fault strike and the current gradient direction.

7. The method for monitoring the deterioration of the mined rock mass structure based on deep learning according to claim 6, characterized in that, The convolution kernel size of the convolution operation is dynamically adjusted according to the rock mass anisotropy parameter.

8. The method for monitoring the deterioration of the mined rock mass structure based on deep learning according to claim 7, characterized in that, Performing cross-stage correlation analysis between the potential failure path heat map and the historical damage feature library using the knowledge distillation mechanism in step S5 includes: extracting the key attention areas of the temporal damage evolution pattern in the historical damage feature library through the teacher model, and extracting the spatio-temporal saliency features of the potential failure path heat map through the student model; Performing spatial position matching on the attention weights of the teacher model and the saliency features of the student model, and the matching degree is calculated as the cosine similarity of the feature vectors at the corresponding positions; The generation of the dynamic weight matrix includes the product of the time decay factor and the spatial correlation weight. The time decay factor decays exponentially according to the time interval between mining stages, and the spatial correlation weight is determined according to the linear combination of the matching degree and the heat value; When there is a fault structure in the mining area, the spatial correlation weight is corrected according to the cosine value of the angle between the fault strike and the heat gradient direction.

9. The method for monitoring the deterioration of the mined rock mass structure based on deep learning according to claim 8, characterized in that, When re-weighting the parameters of the fully connected layer of the damage assessment model according to the dynamic weight matrix in step S6, each element of the dynamic weight matrix is multiplied by the connection weight of the corresponding neuron in the fully connected layer, and the product result is used as the updated connection weight; The updated connection weight is restricted in terms of the change amplitude through the gradient clipping operation, and the clipping threshold is set based on the statistical fluctuation range of the historical weight parameters; The generation of the dynamic monitoring model requires verifying the convergence of the model after parameter update. The convergence criterion is that the loss function value of the model on the validation set decreases by less than the set ratio for multiple consecutive iterations; The output of the rock mass structure deterioration risk assessment result includes the risk level division and spatial position annotation of the mining area. The risk level is mapped to the preset risk interval through the weighted sum result of the dynamic weight matrix, and the spatial position annotation is strictly aligned with the spatial grid coordinates of the mining face.

10. A mining-induced rock mass structure degradation monitoring system based on deep learning, which is used to implement the mining-induced rock mass structure degradation monitoring method according to any one of claims 1-9, characterized in that, Including the following modules: The dynamic acquisition module is used to collect multi-source heterogeneous monitoring data under mining disturbances in real time and construct a dynamic damage increment data set; The feature extraction module is used to extract the newly added damage features in the current mining stage based on the dynamic damage increment data set, and generate a dynamic distribution offset degree based on the similarity measurement between the newly added damage features and the historical damage feature library; A mechanical verification module, which is used to verify the mechanical behavior constraints of newly added damage features when the dynamic distribution offset exceeds the offset threshold, and screen out an effective damage feature subset that conforms to the mechanical laws of rock mass failure; A path modeling module, which is used to perform damage path dependence modeling on the effective damage feature subset, match the damage evolution patterns in the historical damage feature library in the same geological structure area, and generate a potential failure path heat map based on the spatial gradient; A knowledge distillation module, which is used to perform cross-stage correlation analysis on the potential failure path heat map and the historical damage feature library by using the knowledge distillation mechanism, and generate a dynamic weight matrix reflecting the mapping relationship between the mining time sequence and the damage evolution; A dynamic update module, which is used to re-weight the parameters of the fully connected layer of the damage assessment model according to the dynamic weight matrix, generate a dynamic monitoring model and output the rock mass structure deterioration risk assessment result.

Images