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

Abstract

The present invention discloses a method and system for monitoring mining-induced rock structure degradation based on deep learning, which specifically relates to the field of mine safety monitoring technology. The method is used to solve the problems of monitoring lag and misjudgment caused by the inability of existing static models to autonomously adapt to the dynamic evolution of rock damage. A dynamic damage incremental data set is constructed by real-time acquisition of multi-source heterogeneous monitoring data under mining disturbances, and the newly added damage features are extracted and their distribution deviation from historical features is calculated. When the deviation exceeds the limit, a subset of damage features that conforms to mechanical laws is verified and screened based on the rock constitutive equation to block non-physical noise interference. The geological structure evolution pattern is matched through damage path dependency modeling, and a potential damage path heat map is generated in combination with spatial gradients. The knowledge distillation mechanism is used to associate the heat map with the historical feature library across stages to generate a dynamic weight matrix that integrates the mining time series. Finally, the damage assessment model is updated through a parameter reweighting mechanism to achieve adaptive assessment of rock degradation risks.

Description

Technical Field

[0001] The present invention relates to the technical field of mine safety monitoring, and more specifically, to a method and system for monitoring mining-induced rock mass structural degradation based on deep learning. Background Art

[0002] In the field of mine safety monitoring, data-driven rock mass structural degradation assessment methods have gradually replaced traditional manual observation methods. Existing technologies generally use acoustic, seismic, and optical rock mass data collected during historical mining phases to construct training sets. Machine learning models are then used to establish a mapping between rock mass damage characteristics and stability, enabling automated monitoring and early warning. These models rely on data distribution assumptions from the training phase and are then directly deployed in monitoring scenarios across different mining periods.

[0003] Due to the irreversibility and dynamic development characteristics of rock damage evolution, the mechanical behavior of newly formed cracks during the mining process is essentially different from the historical damage pattern, resulting in the inability of static training sets to cover the new feature space of real-time monitoring data. After the model is deployed, the existing methods lack the ability to autonomously adapt to the dynamic degradation laws of rock structure, causing the evaluation results of the monitoring model under the new damage pattern to deviate from the actual physical state, forming a monitoring blind spot, which directly affects the accuracy and timeliness of mine safety warnings. 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 mining-induced rock mass structure degradation based on deep learning to solve the problems raised in the above-mentioned background technology.

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

[0006] The deep learning-based method for monitoring mining-induced rock mass structural degradation includes the following steps:

[0007] S1. Real-time collection of multi-source heterogeneous monitoring data under mining disturbances to construct a dynamic damage increment dataset;

[0008] S2. Extracting new damage features of the current mining stage based on the dynamic damage increment dataset, and generating dynamic distribution deviation based on the similarity measurement between the new damage features and the historical damage feature library;

[0009] S3. When the dynamic distribution deviation exceeds the deviation threshold, the newly added damage features are verified by mechanical behavior constraints to select a subset of effective damage features that conform to the laws of rock mass failure mechanics;

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

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

[0012] S6. Reweight 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 structure degradation risk assessment results.

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

[0014] The energy spectrum data of microseismic events is collected through an underground microseismic monitoring network arranged in a cellular topology. The geophone array covers the boundaries of the goaf and meets the spatial sampling density requirements.

[0015] Borehole imaging data is collected through a borehole television imaging system, and the camera resolution and frame rate match the requirements for observing rock fracture zones;

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

[0017] In a preferred embodiment, in step S2, a feature encoder including a convolutional layer and a fully connected layer is used to extract new damage features from the dynamic damage increment dataset, and the convolution kernel size of the feature encoder is adapted to the spatial distribution characteristics of the rock damage;

[0018] Based on the mean and covariance matrix of each historical feature vector in the historical damage feature library, the Mahalanobis distance of the newly added damage features is calculated as the similarity measure;

[0019] The dynamic distribution deviation is generated by comparing the normalized Mahalanobis distance with the preset similarity threshold, where the normalization coefficient is the inverse of the number of samples in the historical feature library;

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

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

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

[0023] Based on the law of energy conservation during damage evolution, the energy coupling coefficient between microseismic energy release and acoustic wave velocity changes is calculated, and non-physical characteristics of the energy coupling coefficient that deviate from the preset range are eliminated;

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

[0025] When the mining stage enters the periodic pressure period, the friction angle parameters constrained by the stress state are dynamically modified according to the real-time ground stress monitoring data.

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

[0027] A spatial gradient direction field is generated based on the spatiotemporal distribution characteristics of the effective damage feature subset. The gradient direction of each node in the direction field is determined by the difference vector of the damage degree of adjacent nodes.

[0028] The potential damage path heat map is generated by 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 ​​modified 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 anisotropy parameters of the rock mass.

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

[0032] The attention weights of the teacher model are spatially matched with the salient features of the student model, 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 the time decay factor and the spatial correlation weight. The time decay factor decays exponentially according to the interval length of the mining phase, and the spatial correlation weight is determined by 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 modified according to the cosine value of the angle between the fault strike and the thermal gradient direction.

[0035] In a preferred embodiment, when reweighting 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 is used as the updated connection weight;

[0036] The updated connection weights are limited in their variation through gradient clipping, and the clipping threshold is set based on the statistical fluctuation range of the historical weight parameters;

[0037] The generation of a dynamic monitoring model requires verification of the model convergence after parameter updates. The convergence criterion is that the loss function value of the model on the validation set decreases by less than a set ratio for multiple consecutive iterations.

[0038] The output of the rock structure degradation risk assessment results includes the risk level classification and spatial position marking of the mining area. The risk level is mapped to the preset risk range 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 working face.

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

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

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

[0042] The mechanical verification module is used to verify the mechanical behavior constraints of the newly added damage features when the dynamic distribution deviation exceeds the deviation threshold, and to select a subset of effective damage features that conform to the laws of rock mass failure mechanics;

[0043] Path modeling module, which is used to perform damage path dependency modeling on a subset of valid damage signatures, match the damage evolution patterns of the same geological structure area in the historical damage signature library, and generate a potential damage path heat map based on spatial gradients;

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

[0045] The dynamic update module is used to reweight 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 structure degradation risk assessment results.

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

[0047] 1. Through the deep synergy mechanism of dynamic fusion of multi-source monitoring data and physical and mechanical laws, a closed-loop feedback system for rock damage evolution is constructed. A dynamic damage incremental data set is constructed based on real-time mining disturbance data, breaking through the dependence of traditional static training sets on historical data distribution. Through incremental learning, the spatiotemporal evolution characteristics of new cracks are continuously captured, effectively solving the problem of model inaccuracy caused by the irreversibility of rock damage. A mechanical behavior constraint verification mechanism is introduced to screen the constitutive equations of newly added damage characteristics, blocking the interference of non-physical damage modes, ensuring that model updates always conform to the mechanical mechanism of rock fracture, and significantly improving the physical credibility of damage assessment. By coupling geological structure matching with spatial gradient analysis, historical damage patterns are converted into spatiotemporal 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. Through the knowledge distillation mechanism, cross-stage spatiotemporal associations are established, and mining time series information is encoded into a dynamic weight matrix, giving the model the ability to dynamically adapt to the rock degradation process. The parameter reweighting process is combined with gradient clipping and dynamic normalization strategies to achieve precise control of the parameter space while maintaining model stability, forming a dual constraint of data-driven and physical laws, so that the monitoring model can autonomously adapt to the characteristic distribution offset caused by mining disturbances and correct assessment deviations in real time. Through the deep coupling of physical mechanisms and data characteristics, interpretable modeling of the rock damage evolution process is achieved in complex mining environments, providing decision-making support that is both real-time and reliable for mine safety monitoring, and effectively solving the problems of early warning lag and misjudgment caused by model rigidity in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of the method for monitoring mining-induced rock mass structure degradation based on deep learning of the present invention;

[0050] Figure 2 This is a structural schematic diagram of the mining-induced rock structure degradation monitoring system based on deep learning of the present invention. DETAILED DESCRIPTION

[0051] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0052] Example 1: Figure 1 The present invention provides a method for monitoring mining-induced rock mass structural degradation based on deep learning, which includes the following steps:

[0053] S1. Real-time collection of multi-source heterogeneous monitoring data under mining disturbances to construct a dynamic damage increment dataset;

[0054] S2. Extracting new damage features of the current mining stage based on the dynamic damage increment dataset, and generating dynamic distribution deviation based on the similarity measurement between the new damage features and the historical damage feature library;

[0055] S3. When the dynamic distribution deviation exceeds the deviation threshold, the newly added damage features are verified by mechanical behavior constraints to select a subset of effective damage features that conform to the laws of rock mass failure mechanics;

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

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

[0058] S6. Reweight 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 structure degradation risk assessment results.

[0059] S1. Real-time collection of multi-source heterogeneous monitoring data under mining disturbances to construct a dynamic damage increment dataset, including:

[0060] Multi-source heterogeneous monitoring data include 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 rock mass acoustic wave propagation velocity data, the spacing between the acoustic wave transmitter and receiver is dynamically adjusted based on the rock mass integrity. The rock mass integrity coefficient is calculated using the rock quality index (RQD) value of the drilled core. The RQD value is the percentage of the cumulative length of intact core greater than 10 cm in the borehole to the total borehole length. For example, when the RQD value is greater than 75%, the spacing between the transmitter and receiver is set to 50 meters. When the RQD value is less than 75%, the spacing is shortened to 30 meters to accommodate the acoustic wave attenuation characteristics of rock masses with different integrity levels. The acoustic wave frequency range is selected based on the target size of the crack to be detected, calculated as v / (2d), where v is the rock mass longitudinal wave velocity (in meters per second) and d is the average crack opening width (in meters). For example, a lower frequency limit of 10 kHz corresponds to the detection of cracks larger than 5 mm, and an upper frequency limit of 50 kHz corresponds to the detection of cracks of 1 mm.

[0062] The honeycomb topology of the underground microseismic monitoring network is laid out based on the geometric shape and stress distribution characteristics of the goaf. 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 meters in the initial mining stage to cover a large area of ​​low stress area, and is shortened to 20 meters in the periodic pressure stage to focus on the high stress concentration area; the range of the three-component acceleration sensor of the detector node is determined according to the energy of the historical maximum microseismic event, for example, the upper limit of the range is 1.5 times the maximum historical microseismic peak acceleration; the microseismic event triggering threshold is set by statistical analysis of the background noise, for example, taking 3 times the standard deviation of the root mean square value of the background noise acceleration for 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 the maximum value within 24 hours multiplied by 3 as the threshold.

[0063] The camera resolution and light source brightness 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 aperture is less than 75 mm, the resolution is 1920×1080 pixels, and when the aperture is greater than 75 mm, it is increased to 2560×1440 pixels; the light source brightness 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 every meter of drilling depth, and if the dust concentration sensor in the hole detects that the transmittance is lower than 70%, the brightness is increased by an additional 1000 lumens; the calibration scale verification in the fracture width calculation is achieved through the reflective marks preset in the hole, and the mark spacing error 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, and the acoustic data acquisition equipment is synchronized with the master clock through the NTP protocol. A linear regression algorithm is used to compensate for clock deviation. For example, the clock drift rate is calculated with a 5-minute period and dynamically corrected. The timestamp of the borehole imaging data is calculated based on the probe lifting speed and hole depth. The lifting speed error is controlled through encoder feedback closed-loop control. For example, the speed fluctuation range does not exceed ±5% of the set value. The aligned multi-source data are sliced ​​and fused according to the time window. The time window length is set according to the rock damage evolution rate. For example, the window length is 10 minutes when the damage rate is lower than 0.1 mm / h and is shortened to 5 minutes when the damage rate is higher than 0.1 mm / h.

[0065] During the data normalization processing of the dynamic damage increment dataset, the normalized baseline value of the acoustic wave velocity is dynamically adjusted according to the lithology. For example, 5500 m / s is used for granite rock mass, and 4500 m / s is used for sandstone rock mass. The logarithmic transformation base of the microseismic energy spectrum is selected according to the background noise level in the monitoring area. For example, when the noise level is lower than 0.01 m / s², the natural logarithm is used, while when it is higher than 0.01 m / s², the logarithm with base 10 is used to suppress noise interference. The normalization threshold of the borehole fracture width is determined by the historical maximum fracture width statistics. For example, the historical maximum value is recalculated every 50 meters of mining distance.

[0066] The abnormal data processing mechanism includes the following rules for determining acoustic signal anomalies: if the deviation of three consecutive wave velocity measurements of the same transmitter-receiver pair exceeds 15%, it is determined to be an equipment failure; microseismic signal channel anomaly detection uses the adjacent node data correlation analysis method, for example, a channel with a correlation coefficient lower than 0.7 is marked as an abnormal channel; borehole image clarity evaluation is based on the Laplacian gradient operator calculation, for example, a re-sampling mechanism is triggered when the gradient value is lower than a set threshold; the data interpolation method is selected according to the data type, for example, spatial Kriging interpolation is used for acoustic wave velocity, temporal linear interpolation is used for microseismic energy spectrum, and interpolation processing is prohibited for borehole fracture data.

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

[0068] The feature encoder consists of convolutional and fully connected layers. The convolution kernel size of the convolutional layer is set according to the spatial distribution characteristics of rock damage. For example, a 3×5 rectangular convolution kernel is used for horizontal damage extension along the working face, and a 5×3 rectangular convolution kernel is used for vertical damage perpendicular to the rock layer. The convolution kernel stride is set to 1×1 to ensure feature map resolution. The padding mode of the convolutional layer is set to "valid", and convolution operations are performed only within the allowable 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 historical feature library dimension 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 preserve nonlinear damage characteristics. A dropout layer is added after the fully connected layer to prevent overfitting, and the dropout ratio is set to 0.2. The dynamic damage increment dataset is zero-mean normalized before input to 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 4000 m / s and the standard deviation is 500 m / s.

[0069] The mean and covariance matrix of each historical feature vector in the historical damage feature library are calculated during the offline training phase. The mean is calculated as the arithmetic mean of each dimension of the historical feature vector. For example, the mean is calculated for each dimension of 1000 historical feature samples. The covariance matrix is ​​calculated as the mean of the outer product of the historical feature vector and its mean deviation. For example, 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 the threshold, it is truncated to avoid numerical instability; the similarity measure is the inverse of the Mahalanobis distance, and the calculation formula is 1 / (1+Mahalanobis distance) to map the measure value to the range of 0 to 1. For example, when the Mahalanobis distance is 2, the similarity measure value is 0.333.

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

[0071] ;

[0072] in, Represents the covariance matrix of the historical damage feature library, with dimension , is the feature dimension; and for An orthogonal matrix that satisfies ( is the identity matrix); is a diagonal matrix, represents singular values; is the singular value cutoff threshold.

[0073] New damage eigenvector The Mahalanobis distance formula is:

[0074] ;

[0075] in, Represents the newly added damage feature vector Mahalanobis distance; Represents the current newly added damage feature vector, with dimension ; Represents the mean vector of the historical feature library, with dimension ; Represents the inverse of the covariance matrix, computed using the modified singular value decomposition.

[0076] In the normalization processing of the dynamic distribution deviation, the normalization coefficient is the inverse of the number of samples in the historical damage signature library. For example, when the sample size 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 deviation, if the normalized Mahalanobis distance exceeds the similarity threshold, it is determined that there is significant distribution deviation, and the deviation 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 deviation value is 0.8 / 0.7≈1.14, but after truncation, 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 proportion of the microseismic energy spectrum in the frequency band above 100 Hz increases by more than 30%, it is determined to be a synchronous enhancement mode; the temporal correlation weighting factor is calculated by linear interpolation based on the advancement rate of the mining working face. For example, when the advancement 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 weighted similarity measurement calculation formula 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 calculated result is 1.2, which is truncated to 1.0.

[0078] The feature encoder training process is completed offline. The training data is all samples in the historical damage feature library. The loss function uses the mean square error function, and the optimizer selects the Adam algorithm. The initial learning rate is set to 0.001, and it decays to 90% of the original value every 50 training cycles. The training termination condition is that the validation set loss decreases by less than 1% for 10 consecutive cycles. The final model parameters are saved as the initial weights of the feature encoder. The training hardware environment must support CUDA parallel computing, such as using the NVIDIA Tesla V100 GPU to accelerate the training process. The software dependencies include TensorFlow 2.4 and above.

[0079] The exception handling mechanism includes the detection and repair of covariance matrix singularities, and when the matrix condition number is greater than 10 6 When it is judged to be close to singularity, the diagonal loading method is used to add 10 -5 The unit matrix of the order of magnitude is used to enhance stability; if a negative value appears during the Mahalanobis distance calculation, a data integrity check is immediately triggered, and the feature vector of the previous time series is recalculated; a sliding average filter is performed before the dynamic distribution offset is output, and the window length is 5 time steps to suppress instantaneous noise interference; if the input data dimension does not match the feature encoder inference stage, the dimension padding or truncation is automatically triggered, and the padding value is the global mean of the corresponding feature.

[0080] Boundary condition processing includes strategies for dealing with extreme input data. For example, when the sudden drop in acoustic wave velocity exceeds 50%, it is determined to be a sensor failure and switch to the backup data channel. When the microseismic energy spectrum shows full-band saturation, dynamic gain adjustment is initiated to temporarily expand the sensor range to 2 times. The covariance matrix update cycle is set to recalculate once every 100 new feature samples to ensure the timeliness of the matrix. The upper limit of the historical damage feature library is set to 10,000 samples. After the upper limit is reached, the old samples are replaced on a first-in-first-out basis.

[0081] S3. When the dynamic distribution deviation exceeds the deviation threshold, the newly added damage features are verified by mechanical behavior constraints to select a subset of effective damage features that conform to the laws of rock mass failure mechanics, including:

[0082] When the dynamic distribution deviation exceeds the deviation threshold, the mechanical behavior constraint verification of the newly added damage characteristics is carried out based on the rock mass constitutive equation. First, the Mohr-Coulomb criterion is used to verify whether the stress state corresponding to the damage characteristic is within the shear strength envelope of the rock mass. The cohesion parameter and friction angle parameter of the Mohr-Coulomb criterion are dynamically set according to the rock mass lithology category. 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 a laboratory core uniaxial compression test with a test loading rate of 0.5 mm / min until the rock sample is destroyed, and the peak stress is recorded as Cohesion value; friction angle parameter is determined by direct shear test. The positive stress of direct shear test is applied in stages. The shear displacement and shear stress are recorded under each level of positive stress. The shear stress-displacement curve is drawn, and the angle corresponding to the peak shear stress of the curve is taken as the friction angle; the shear strength envelope is calculated by the relationship between 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 feature, and the shear stress is half of the difference between the maximum principal stress and the minimum principal stress; during verification, if the calculated shear stress exceeds the theoretical value of the envelope corresponding to the current normal stress, it is determined to be a feature that violates the law of rock shear strength and is eliminated. At the same time, an abnormal feature report is generated, which includes the abnormal stress value, verification time and associated sensor number.

[0083] The determination of the dynamic distribution deviation threshold needs to be combined with statistical analysis of historical mining data. The specific implementation includes: collecting at least three months of historical dynamic distribution deviation data in the initial stage, and the data should cover different mining intensities (such as daily advancement rate of 1-5 meters) and geological conditions (such as fault zones and intact rock formations); performing a Kolmogorov-Smirnov test on the data distribution. If the test P value is greater than 0.05, it is determined to be normally distributed, and the threshold is set as the mean plus three times the standard deviation; otherwise, it is set according to the 95th percentile; for example, if the historical deviation data of a mining area is found to be non-normally distributed after inspection, and the 95th percentile after sorting is 0.85, the initial threshold is set to 0.85; the threshold is updated once a month, and the latest 30 days of deviation data are included in the update to recalculate the quantile or mean standard deviation; when the geological conditions of the mining area change significantly (such as the exposure of a large fault), the threshold is immediately triggered to be recalculated and manually reviewed and confirmed.

[0084] The energy coupling coefficient between microseismic energy release and acoustic wave velocity changes is calculated based on the law of energy conservation during damage evolution. The calculation logic of the energy coupling coefficient is the ratio of the energy released by the microseismic event to the energy consumed by the acoustic wave velocity change. The microseismic energy is obtained by integrating the area of ​​the microseismic event energy spectrum within the time window. The integration interval is 10 seconds before and 20 seconds after the event trigger, and the frequency resolution is 1 Hz. The energy consumed by the acoustic wave velocity change is calculated by multiplying the velocity change rate by the rock density. The velocity change rate is obtained by dividing the velocity difference between adjacent time windows by the time interval. The rock density is determined by laboratory measurements of drill core. The preset interval is determined by statistically calculating the historical data from normal mining phases. The statistical method is to calculate the mean of the historical energy coupling coefficient plus or minus two standard deviations. 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 density is kilograms per cubic meter. All unit conversion coefficients are uniformly calibrated before data processing.

[0085] The screening criteria for the effective damage feature subset are the feature set that satisfies both the stress state constraint and the energy conservation constraint. The implementation method is to perform stress state verification and energy coupling verification on each newly added damage feature in sequence, and only retain the features that pass both verifications; the stress state verification is recorded through an independent state mark bit, where the mark bit is 1 for passed verification and 0 for failed verification; the energy coupling verification is recorded through another independent mark bit, and the mark bit generation rule is that the energy coupling coefficient is set to 1 when it is within the preset range, otherwise it is set to 0; finally, a comprehensive verification mark bit is generated through a logical AND operation, and the comprehensive mark bit is set to 1 only when both mark bits are 1; the feature set with the comprehensive mark bit 1 is transmitted to the subsequent processing link, and the features with the mark bit 0 are stored in the abnormal feature library, which is stored by timestamp and spatial position index to support rapid positioning during manual review.

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

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

[0088] ;

[0089] in, represents the corrected friction angle (unit: degree); represents the initial friction angle (unit: degree), which is determined by laboratory direct shear test; It represents the correction coefficient (unit: degree / unit ground stress ratio), which is determined by linear regression analysis of ground stress ratio and friction angle changes in rock mechanics tests, and its value range is ; represents the current ground stress ratio, is the horizontal ground stress (unit: MPa), is the vertical ground stress (unit: MPa), which is measured in real time by a borehole stress gauge; Indicates the benchmark value of the ground stress ratio, which is calculated by the average of the historical ground stress ratio data during the mine development stage.

[0090] The calculation formula is:

[0091] ;

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

[0093] Corrected friction angle It is updated every 2 hours. If the ground stress data is missing, the corrected value of the previous period will be used and a data integrity alarm will be triggered.

[0094] The anomaly handling mechanism includes extreme stress handling rules in stress state verification. When the maximum principal stress exceeds 80% of the rock mass's uniaxial compressive strength, the manual review process is automatically triggered and the automated screening process is suspended. In the energy coupling coefficient calculation, if the acoustic wave velocity change rate is zero, the calculation of the current time window is skipped and the data is marked as invalid. Invalid data segments are marked as "manual verification required" in the database. When the ground stress data is missing during the dynamic correction of the friction angle, the sliding average method is used to interpolate and supplement the data based on the previous 6 hours of data. The interpolation weights are distributed according to time proximity and distance to [0.4, 0.3, 0.2, 0.1]. When all abnormal events are triggered, a timestamp log record is generated. The log contains fields such as the anomaly type, associated feature index, handling measures, and operator signature. The log files are archived and stored encrypted on a daily basis.

[0095] Boundary condition coverage includes adaptive adjustment rules for the energy coupling coefficient interval. When the number of valid features in five consecutive time windows is less than 50% of the historical average, the interval range is expanded according to the preset step size, which is 10% of the original interval width. The maximum expansion amplitude does not exceed 50% of the original interval. The normal stress value range in stress state verification is limited to 0 to 80% of the uniaxial compressive strength of the rock mass. Features out of the range are directly discarded and recorded as "out-of-limit data". The friction angle correction value is forcibly limited to the allowable range of the rock mass physical properties. 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 truncation according to the maximum value. The truncation operation is recorded in the log and triggers a system notification.

[0096] S4. Perform damage path dependency modeling on the valid damage signature subset, match the damage evolution pattern of the same geological structure area in the historical damage signature library, and generate a potential damage path heat map based on spatial gradients, including:

[0097] When performing damage path dependency modeling on the effective damage feature subset, the rock fracture propagation direction and rate parameters are extracted according to the damage evolution pattern of the same geological structure area in the historical damage feature library to construct a path similarity matrix; the matching method for the same geological structure area is as follows: the geological structure parameters of the current mining area, including fault strike, dip, and rock layer thickness, are calculated with the geological structure parameters recorded in the historical feature library. The top five historical areas with the smallest Euclidean distance are selected as reference patterns; the fracture propagation direction parameter is determined by the average azimuth of the fracture trace in the historical area, and the azimuth of the fracture trace is determined by the fracture contour imaged by the borehole television. The angle between the fitting line and the north direction is measured; the rate parameter is calculated by the rate of change of the crack length over time. The rate of change is the ratio of the crack length difference to the time interval. The time interval is dynamically adjusted according to the mining stage. The initial mining stage is 24 hours, and the periodic pressure stage is shortened to 6 hours. Each element of the path similarity matrix represents the similarity between the current damage feature and the historical pattern in direction and rate. The direction similarity is calculated as the cosine value of the angle between the current crack direction and the historical crack direction. The rate similarity is calculated as the ratio of the current rate to the historical rate and is limited to between 0.1 and 10. The final similarity is the product of the direction cosine value and the rate ratio.

[0098] A spatial gradient direction field is generated based on the spatiotemporal distribution characteristics of a subset of effective damage features. The gradient direction of each node in the spatial gradient direction field is determined by the difference vector of the damage levels of adjacent nodes. The spatial grid division is set according to the advancement direction of the mining working face, with a grid cell size of 5 meters × 5 meters. Adjacent nodes are defined as six nodes directly connected in space: up, down, left, right, front, and back. The difference vector is calculated as the difference vector between the damage level of the current node and the damage level of the adjacent node. 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 level difference. During the generation of the gradient direction field, if the data of adjacent nodes is missing, the inverse distance weighted interpolation method is used to supplement the data. The interpolation weight is inversely proportional to the square of the distance between the 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, retaining the direction field of the previous period as a reference benchmark during the update. If the current direction field deviates from the reference field by more than 30 degrees, a direction anomaly alarm is triggered and manual verification is initiated.

[0099] The potential damage path heat map is generated by convolving the path similarity matrix with the spatial gradient direction field. The convolution kernel size is dynamically adjusted according to the rock mass anisotropy parameters. The rock mass anisotropy parameters are obtained through wave velocity anisotropy testing of laboratory rock cores. During the test, sound waves are applied along the axial, tangential, and normal directions of the rock core and the wave velocity is recorded. The anisotropy difference ratio is calculated as (maximum wave velocity - minimum wave velocity) / average wave velocity × 100%. The lateral size of the convolution kernel (along the mining direction) is set as a linear function of the anisotropy difference ratio. For example, when the difference ratio is 20%, the lateral size is 7 × 3, and the longitudinal size is 3 × 7.

[0100] The convolution operation uses zero padding mode to keep the output heat map size unchanged. The zero padding convolution formula is:

[0101] ;

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

[0103] The convolution result is normalized to the range of 0 to 1 using the Sigmoid function. The normalization formula is:

[0104] ;

[0105] in is the original value of the convolution result; Represents the normalized thermal value, with a value range of [0,1]; is the base of natural logarithms (Euler's number), ; When the area is judged as a high-risk area, an early warning signal is triggered, and the threshold of 0.8 is set based on historical accident data statistics.

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

[0107] When a fault structure is exposed at the mining face, the weight coefficient of the path similarity matrix is ​​modified based on the cosine value of the angle between the fault strike and the current gradient direction. The fault strike is determined by jointly interpreting borehole television imaging and geological radar data. The interpretation method is to extract the angle between the extension direction of the fault interface and the 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 two vector moduli, and the calculated result is limited to between -1 and 1. The weight coefficient correction formula is the original weight multiplied by (1 + cosine value). The corrected weight coefficient is limited to the range of 0 to 1 using a truncation function. For example, if the corrected calculation result is 1.2, it is 1.0, and if the calculation result is -0.3, it is 0. The trigger condition for weight coefficient update is that the working face advances to within 50 meters of the fault influence zone, and the weight coefficient is recalculated every 10 meters. If the weight coefficient fluctuates by more than 50% between two consecutive calculations, the automatic correction is suspended and a manual intervention request is generated.

[0108] The exception handling mechanism includes the singular value processing of the path similarity matrix. When the matrix condition number is greater than 10 6 When adding 10 -5 Regularization is performed on the unit matrix of the order of magnitude; if more than 50% of the adjacent node data are missing during the interpolation of the spatial gradient direction field, it is determined to be an invalid area and the thermal map generation is suspended. The calculation is 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 is >30%), the preset maximum convolution kernel size of 11×5 is used, and the exceeding limit event is recorded in the log; after the thermal map is generated, if the thermal value of the same area in three consecutive time windows continues to exceed 0.9, a red alert is triggered and the manual verification process is started. The verification results are fed back to the historical damage feature library to update the damage evolution model.

[0109] The boundary conditions include similarity range limits for the path similarity matrix. When the rate ratio exceeds 10, it is forcibly truncated to 10 and marked as abnormal data. When the modulus length of the node damage degree difference vector of the spatial gradient direction field exceeds twice the historical maximum value, it is judged as a sensor anomaly and the backup data source is activated.

[0110] S5. Use 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 mining time sequence and damage evolution, including:

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

[0112] When spatially matching the attention weights of the teacher model with the salient features of the student model, the grid division of the spatial matching is consistent with the spatial resolution of the potential damage path heat map. The grid unit size is 5 meters × 5 meters, which is aligned with the sensor layout grid of the mining working face. The matching degree is calculated as the cosine similarity between the attention weight vector of the teacher model and the salient feature vector of the student model in the corresponding grid unit. The cosine similarity calculation formula is the dot product of the two vectors divided by the product of the vector modulus. If there is a fault structure in the grid unit, the matching degree needs to be multiplied by the fault influence factor. 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 rock mechanics experiments.

[0113] The dynamic weight matrix is ​​generated by multiplying the time decay factor by the spatial correlation weight. The time decay factor is calculated exponentially based on the duration of the mining phase interval. Its base is the base e of the natural logarithm, and the exponential component is the product of the negative decay coefficient λ and the number of days between phases, Δt. The decay coefficient λ is determined by fitting a curve of the damage evolution rate and the time decay effect in historical mining data. Specifically, the relationship between the damage rate and time decay is fitted using the least squares method. For example, when the damage rate is 0.1 mm per day, λ is set to 0.01; when the damage rate increases to 0.5 mm per day, λ is adjusted to 0.05. The number of days between phases, Δt, is the time difference between the current mining phase and the historical reference phase, calculated in calendar days. The range of λ values ​​is based on the results of regression analysis of historical data to ensure the model's adaptability to the mining process. The spatial correlation weight is determined by a linear combination of the matching degree and the thermal value. The linear combination coefficient is obtained through multivariate 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 extension probability. The regression results show that the matching degree weight accounts for 60% and the thermal value weight accounts for 40%. Each element value of the dynamic weight matrix is ​​normalized to the range of 0 to 1 using the Sigmoid function. When it exceeds 1, it is truncated to 1, and when it is less than 0, it is set to 0. The truncation event is recorded in the exception log.

[0114] When fault structures exist in the mining area, the spatial correlation weight is modified based on the cosine of the angle between the fault strike and the thermal gradient direction. The fault strike is determined by jointly interpreting the joint statistics of the drill core and the geological radar scanning data. The interpretation method is to extract the angle between the extension direction of the fault interface and the north direction. The interpretation error is controlled within ±3 degrees by taking the average of three measurements. The cosine of the angle 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 to between -1 and 1. The modified spatial correlation weight is the original weight multiplied by (1 + cosine value). For example, if the original weight is 0.6 and the cosine value is 0.5, the modified weight is 0.6 × 1.5 = 0.9. If the modified weight exceeds 1, it is set to 1, and if it is less than 0, it is set to 0. The modification operation triggers a log record, which contains the fault number, modification time, operator identification, and the weight values ​​before and after modification.

[0115] The exception handling mechanism includes alignment and verification of the feature dimensions of 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 truncated part retains the first N-dimensional features (N is the output dimension of the student model). During the dynamic weight matrix generation process, if more than 50% of the grid cell weights are detected to be 0, it is judged as a data anomaly and the weight matrix of the previous time series is rolled back, and the data quality check process is triggered at the same time. If the interpretation data is missing when calculating the fault impact factor, the correction mechanism is suspended and the uncorrected weights are used to continue the calculation. The missing data window is marked as "to be supplemented". Hardware dependencies include NVIDIA GPU (such as Tesla V100) to accelerate matrix operations. CUDA 11.3 and above and the corresponding driver must be configured. The software environment must install the PyTorch 1.9 framework and load the pre-trained self-attention model parameters. The parameter file format is the .pt format officially supported by PyTorch.

[0116] The boundary conditions include a maximum interval of 365 days for the time attenuation factor. When the interval exceeds 365 days, the attenuation coefficient λ is reset to the default value of 0.02 to prevent numerical overflow. The linear combination coefficient of the spatial correlation weight is recalibrated once a month. The calibration data source is the valid feature samples of the last three months. If the number of samples is insufficient, the coefficient of the previous month will be used.

[0117] S6. Reweight 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 degradation risk assessment results, including:

[0118] When reweighting the parameters of the fully connected layer of the damage assessment model based on the dynamic weight matrix, each element of the dynamic weight matrix is ​​multiplied element-by-element by the connection weight of the corresponding neuron in the fully connected layer, and the product is used as the updated connection weight. The dimension of the dynamic weight matrix is ​​consistent with the number of input neurons in 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, ensuring that the weight of each neuron connection is dynamically adjusted. The parameter reweighting 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 limited in their variation 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 range of the historical weights during the training phase and set the clipping threshold to a fixed multiple of the average fluctuation range. For example, when the historical average fluctuation range is 0.2, the clipping threshold is 0.3. Gradient clipping adopts a global norm clipping strategy to limit the weight update step size to no more than the threshold to prevent gradient explosion or vanishing.

[0119] The generation of a dynamic monitoring model requires verification of the model convergence after parameter update. The convergence criterion is that the loss function value of the model on the validation set decreases by less than a set ratio for multiple consecutive iterations. The validation set is composed of data that did not participate in training in the historical damage feature library, and the data is divided into a training set and a validation set according to a preset ratio, for example, 8:2. The decrease in the loss function value is calculated as the relative change rate of the loss value of two adjacent iterations. The set ratio is dynamically adjusted according to the complexity of the model. 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 decrease in the loss value for multiple consecutive iterations is less than the set ratio, the model is judged to have converged and the parameter update is locked. If the convergence criterion is not met, the model will fall back to the previous weight parameter and trigger the learning rate adjustment mechanism. The learning rate adjustment range is dynamically set according to the trend of loss value changes.

[0120] The output of the rock structure degradation risk assessment results includes the risk level classification and spatial position marking of the mining area. The risk level is mapped to the preset risk interval through the weighted summation result of the dynamic weight matrix; the weighted summation is calculated as the weight value corresponding to each spatial position in the dynamic weight matrix multiplied by the sum of its associated damage characteristic value, and the summation result is normalized to the preset interval, such as 0 to 100, through linear transformation; the preset rules for risk intervals are: the low risk interval corresponds to the normalized value of 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 marking is strictly aligned with the spatial grid coordinates of the mining working face. The grid coordinates are generated according to the geographic registration information of the drilling imaging data. The center point coordinates of each grid unit are calibrated by high-precision GPS positioning equipment, and the positioning error is controlled within the allowable range, such as ±0.1 meters; the risk marking results are superimposed on the three-dimensional mine model to support interactive visual query.

[0121] The exception handling mechanism includes an automatic rollback strategy when parameter updates fail. If the model convergence verification fails and the learning rate adjustment exceeds the preset number of times, it automatically rolls back to the initial weights and generates a manual intervention request. If it is detected that the weight change exceeds the preset ratio during the gradient clipping process, it is determined to be a gradient abnormality event, the parameter update is suspended, and the model initialization process is restarted. Spatial consistency verification is required before the risk assessment results are output. If the risk levels of adjacent grid cells span multiple intervals, spatial smoothing filtering is triggered, and the filter window size is dynamically adjusted according to the grid resolution. For example, a 3×3 filter window is used for a 5-meter resolution grid. Hardware dependencies include graphics processors that support parallel computing, such as NVIDIA Tesla series GPUs. Software dependencies include deep learning frameworks and geographic information processing libraries, such as TensorFlow 2.6 and GeoPandas 0.10.

[0122] The boundary conditions cover the numerical range constraints of the dynamic weight matrix. The updated weight parameters are forced to be limited between -1 and 1. The values ​​out of the range are compressed to this interval through a nonlinear 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 the model is marked as invalid.

[0123] This embodiment constructs a closed-loop feedback system for rock damage evolution through the dynamic fusion of multi-source heterogeneous monitoring data and the deep embedding of physical and mechanical constraints. During the data acquisition phase, a spatial topology optimization strategy for cross-hole acoustic and microseismic monitoring is employed to overcome the shortcomings of traditional sensor placement in responding poorly to the anisotropic nature of geological structures. A rock constitutive equation verification mechanism is introduced to filter damage signatures that conform to mechanical laws in real time, preventing interference from noisy data on model updates. Damage path dependency modeling creatively couples historical geological structure matching with spatial gradient analysis, transforming static empirical data into spatiotemporal constraints for dynamic evolution path prediction. A knowledge distillation mechanism integrates temporal decay and spatial correlation weights in cross-stage correlation analysis, breaking away from the traditional static allocation of weight matrices. The parameter reweighting process achieves a balance between stability and adaptability in model parameter updates through the synergistic effects of gradient clipping and dynamic normalization. All technical components form an integrated whole: data acquisition provides physically realistic input for feature extraction, mechanical verification ensures the credibility of the evolution model, path modeling imparts geological significance to spatiotemporal predictions, knowledge distillation establishes a cross-stage knowledge transfer channel, and dynamic updates ensure the model's continuous adaptation to the mining process. This multi-dimensional constraint coupling and dynamic feedback mechanism 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] Example 2: Figure 2 The structural diagram of the mining-induced rock mass structure degradation monitoring system based on deep learning of the present invention is given. The mining-induced rock mass structure degradation monitoring system based on deep learning includes the following modules:

[0125] Dynamic acquisition module, used to collect multi-source heterogeneous monitoring data under mining disturbance in real time and construct dynamic damage increment data set;

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

[0127] The mechanical verification module is used to verify the mechanical behavior constraints of the newly added damage features when the dynamic distribution deviation exceeds the deviation threshold, and to select a subset of effective damage features that conform to the laws of rock mass failure mechanics;

[0128] Path modeling module, which is used to perform damage path dependency modeling on a subset of valid damage signatures, match the damage evolution patterns of the same geological structure area in the historical damage signature library, and generate a potential damage path heat map based on spatial gradients;

[0129] The knowledge distillation module is used to analyze the cross-stage correlation between the potential damage path heat map and the historical damage feature library using the knowledge distillation mechanism, and generate a dynamic weight matrix reflecting the mapping relationship between mining time sequence and damage evolution;

[0130] The dynamic update module is used to reweight 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 structure degradation risk assessment results.

[0131] The calculations involved in the embodiments are all dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to actual conditions.

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

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

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

[0135] In the several embodiments provided in this 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 schematic. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as 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 mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0136] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0137] Finally: The above description is only a preferred embodiment of the present invention and is 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 in the scope of protection of the present invention.

Claims

1. A method for monitoring mining-induced rock mass structure degradation based on deep learning, characterized in that: The steps include: S1. Real-time collection of multi-source heterogeneous monitoring data under mining disturbances to construct a dynamic damage increment dataset; S2. Extracting new damage features of the current mining stage based on the dynamic damage increment dataset, and generating dynamic distribution deviation based on the similarity measurement between the new damage features and the historical damage feature library; S3. When the dynamic distribution deviation exceeds the deviation threshold, the newly added damage features are verified by mechanical behavior constraints to select a subset of effective damage features that conform to the laws of rock mass failure mechanics; S4. Perform damage path dependency modeling on the valid damage feature subset, match the damage evolution pattern of the same geological structure area in the historical damage feature library, and generate a potential damage path heat map based on the spatial gradient; S5. Use 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 mining time sequence and damage evolution, including: The teacher model extracts the key attention areas of the temporal damage evolution pattern in the historical damage feature library, and the student model extracts the spatiotemporal saliency features of the potential damage path heat map; The attention weights of the teacher model are spatially matched with the salient 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 interval length of the mining phase, and the spatial correlation weight is determined by the linear combination of the matching degree and the thermal value. When there is a fault structure in the mining area, the spatial correlation weight is modified according to the cosine value of the angle between the fault strike and the thermal gradient direction; S6. Reweight 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 structure degradation risk assessment results.

2. The method for monitoring mining-induced rock mass structure degradation based on deep learning according to claim 1, characterized in that: In step S1, rock mass acoustic wave propagation velocity data is collected by a cross-hole acoustic wave testing device, where the acoustic wave transmitter and receiver are spaced apart along the strike of the mining face and the frequency range is adapted to the scale of the rock mass fractures; The energy spectrum data of microseismic events is collected through an underground microseismic monitoring network arranged in a cellular topology. The geophone array covers the boundaries of the goaf and meets the spatial sampling density requirements. Borehole imaging data is collected through a borehole television imaging system, and the camera resolution and frame rate match the requirements for observing rock fracture zones; The rock mass acoustic wave propagation velocity data, microseismic event energy spectrum data and borehole imaging data are fused into a dynamic damage increment dataset after time stamp alignment.

3. The method for monitoring mining-induced rock mass structure degradation based on deep learning according to claim 2, characterized in that: In step S2, a feature encoder including convolutional layers and fully connected layers is used to extract new damage features from the dynamic damage increment dataset. The convolution kernel size of the feature encoder is adapted to the spatial distribution characteristics of rock damage. Based on the mean and covariance matrix of each historical feature vector in the historical damage feature library, the Mahalanobis distance of the newly added damage features is calculated as the similarity measure; The dynamic distribution deviation is generated by comparing the normalized Mahalanobis distance with the preset similarity threshold, where the normalization coefficient is the inverse of the number of samples in the historical feature library; When the newly added damage features include a sudden drop in acoustic wave velocity and a synchronous enhancement pattern of the high-frequency components of the microseismic energy spectrum, a time-series correlation weighting factor is introduced into the similarity measurement calculation.

4. The method for monitoring mining-induced rock mass structure degradation based on deep learning according to claim 3, characterized in that: The value of the time series correlation weighting factor is positively correlated with the mining advancement rate.

5. The method for monitoring mining-induced rock mass structure degradation based on deep learning according to claim 4, characterized in that: In step S3, when the dynamic distribution deviation exceeds the deviation threshold, the newly added damage feature is verified for mechanical behavior constraints based on the rock mass constitutive equation, including: verifying whether the stress state corresponding to the damage feature is within the rock mass shear strength envelope using the Mohr-Coulomb criterion, and eliminating abnormal features that exceed the strength envelope; Based on the law of energy conservation during damage evolution, the energy coupling coefficient between microseismic energy release and acoustic wave velocity changes is calculated, and non-physical characteristics of the energy coupling coefficient that deviate from the preset range are eliminated; The screening criteria for the effective damage feature subset is the feature set that satisfies both the stress state constraint and the energy conservation constraint; When the mining stage enters the periodic pressure period, the friction angle parameters constrained by the stress state are dynamically modified according to the real-time ground stress monitoring data.

6. The method for monitoring mining-induced rock mass structure degradation based on deep learning according to claim 5, characterized in that: In step S4, damage path dependency modeling is performed on the effective damage feature subset, including: extracting rock mass crack propagation direction and rate parameters to construct a path similarity matrix based on the damage evolution pattern of the same geological structure area in the historical damage feature library; A spatial gradient direction field is generated based on the spatiotemporal distribution characteristics of the effective damage feature subset. The gradient direction of each node in the direction field is determined by the difference vector of the damage degree of adjacent nodes. The potential damage path heat map is generated by 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 ​​modified according to the cosine value of the angle between the fault strike and the current gradient direction.

7. The method for monitoring mining-induced rock mass structure degradation 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 anisotropy parameters of the rock mass.

8. The method for monitoring mining-induced rock mass structure degradation based on deep learning according to claim 7, characterized in that: In step S6, when reweighting 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 is used as the updated connection weight; The updated connection weights are limited in their variation through gradient clipping, and the clipping threshold is set based on the statistical fluctuation range of the historical weight parameters; The generation of a dynamic monitoring model requires verification of the model convergence after parameter updates. The convergence criterion is that the loss function value of the model on the validation set decreases by less than a set ratio for multiple consecutive iterations. The output of the rock structure degradation risk assessment results includes the risk level classification and spatial position marking of the mining area. The risk level is mapped to the preset risk range 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 working face.

9. A system for monitoring structural degradation of mining-induced rock mass based on deep learning, used to implement the method for monitoring structural degradation of mining-induced rock mass based on deep learning according to any one of claims 1 to 8, characterized in that: Includes the following modules: Dynamic acquisition module, used to collect multi-source heterogeneous monitoring data under mining disturbance in real time and construct dynamic damage increment data set; A feature extraction module is used to extract the new damage features of the current mining stage based on the dynamic damage increment dataset, and generate a dynamic distribution offset based on the similarity measurement between the new damage features and the historical damage feature library; The mechanical verification module is used to verify the mechanical behavior constraints of the newly added damage features when the dynamic distribution deviation exceeds the deviation threshold, and to select a subset of effective damage features that conform to the laws of rock mass failure mechanics; Path modeling module, which is used to perform damage path dependency modeling on a subset of valid damage signatures, match the damage evolution patterns of the same geological structure area in the historical damage signature library, and generate a potential damage path heat map based on spatial gradients; The knowledge distillation module is used to analyze the cross-stage correlation between the potential damage path heat map and the historical damage feature library using the knowledge distillation mechanism, and generate a dynamic weight matrix reflecting the mapping relationship between mining time sequence and damage evolution; The dynamic update module is used to reweight 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 structure degradation risk assessment results.

Images