Mine disaster prediction and risk assessment method based on AI

Abstract

The invention relates to the field of mine disaster prediction, and discloses an AI-based mine disaster prediction and risk assessment method, which comprises the steps of deploying a multi-source sensor network which comprises a microseismic monitoring sensor, a gas concentration sensor, a stress sensor and a hydrological sensor, the data acquisition module is used for acquiring micro-seismic waveform data, gas concentration time sequence data, stress data and hydrological parameter data of a mining working face, a roadway and a geological structure area; and performing space-time alignment processing on the microseismic waveform data, the gas concentration time sequence data, the stress data and the hydrological parameter data to generate a space-time aligned heterogeneous data set, and performing encryption feature interaction with an external system through a federated learning framework. According to the method, a multi-modal deep learning model is constructed, microseismic waveforms, gas concentration time sequences and geologic structure space correlation characteristics are extracted, rock mechanics equation constraint conditions are injected in model training, and the recognition capacity of rock stratum fracture precursor signals and the disaster chain coupling process is remarkably enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of mine disaster prediction, and specifically provides an AI-based mine disaster prediction and risk assessment method. Background Technique

[0002] Mine disaster prediction and risk assessment are the core links of mine safety production, involving the real-time collection and analysis of multi-source heterogeneous data such as microseismic waveforms, gas concentrations, stresses, and hydrogeological parameters in working faces, roadways, and geological structure areas. With the advancement of intelligent mining technology in mines, the types of disasters show a trend of diversification and coupling, posing higher requirements for the early identification of multi-disaster coupled evolution, the prediction of propagation paths, and the accuracy of dynamic prevention and control strategies. Currently, the industry urgently needs to break through key technologies such as the efficient integration of multi-dimensional monitoring data, the modeling and analysis of disaster evolution laws under complex geological conditions, and the adaptive optimization of early warning response mechanisms, so as to build a full-process intelligent prevention and control system covering "data collection - risk assessment - emergency response".

[0003] In the prior art, for multi-source heterogeneous data such as microseismic waveforms, gas concentration time series, stresses, and hydrogeological parameters, there is a lack of a differential feature extraction and deep fusion mechanism for different data characteristics, and physical mechanisms such as rock mechanics equations are not incorporated into model training, resulting in insufficient fitting of the internal physical laws of disaster evolution under complex geological conditions, and relatively low identification accuracy of precursor signals of rock layer fractures and prediction accuracy of the coupling process of disaster chains. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides an AI-based mine disaster prediction and risk assessment method, which solves the problems in the prior art that for multi-source heterogeneous data such as microseismic waveforms, gas concentration time series, stresses, and hydrogeological parameters, there is a lack of a differential feature extraction and deep fusion mechanism for different data characteristics, and physical mechanisms such as rock mechanics equations are not incorporated into model training, resulting in insufficient fitting of the internal physical laws of disaster evolution under complex geological conditions, and relatively low identification accuracy of precursor signals of rock layer fractures and prediction accuracy of the coupling process of disaster chains.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions:

[0006] An AI-based mine disaster prediction and risk assessment method includes:

[0007] Deploy a multi-source sensor network, where the sensor network includes microseismic monitoring sensors, gas concentration sensors, stress sensors, and hydrogeological sensors, and is used to collect microseismic waveform data, gas concentration time series data, stress data, and hydrogeological parameter data in working faces, roadways, and geological structure areas;

[0008] The microseismic waveform data, gas concentration time series data, stress data and hydrological parameter data are spatiotemporally aligned to generate a spatiotemporally aligned heterogeneous dataset, and encrypted feature interaction with external systems is performed through a federated learning framework.

[0009] A multimodal deep learning model is constructed based on the heterogeneous dataset. The multimodal deep learning model extracts microseismic waveform features through a one-dimensional convolutional neural network, extracts gas concentration temporal features through a bidirectional long short-term memory network, models the spatial correlation features of geological structures through a graph attention network, and injects rock mechanics equation constraints during the model training process.

[0010] The digital twin simulation module is invoked to generate disaster chain composite data, which includes evolution path label data of rock burst triggering surrounding rock fracture propagation, fracture conduction leading to abnormal gas outburst, and abrupt changes in hydrological parameters.

[0011] Energy mutation detection is performed on real-time collected microseismic waveform data by edge computing devices. Events with energy mutations exceeding a preset threshold are identified as precursor signals of rock strata rupture. At the same time, the cloud server calculates the probability of disaster chain propagation based on the multimodal deep learning model.

[0012] When the probability of the disaster chain propagation exceeds a preset threshold, a multi-level emergency control command is triggered. The multi-level emergency control command includes a power shutdown command, an emergency ventilation parameter adjustment command, and a grouting reinforcement parameter push command.

[0013] Based on the comparison results between historical early warning records and actual disaster data, the feature weights and early warning thresholds of the multimodal deep learning model are dynamically adjusted through a reinforcement learning mechanism. Cluster analysis is performed on historical disaster data to generate risk heat maps, and the priority of risk areas is dynamically updated in conjunction with real-time mining progress to complete risk assessment and optimization of prevention and control strategies.

[0014] Preferably, the spatiotemporal alignment process includes:

[0015] A dynamic time warping algorithm was used to match the time axis of microseismic waveform data, gas concentration time series data, stress data and hydrological parameter data with different sampling frequencies.

[0016] Energy accumulation calculations were performed on the wavelet packet transform-denoised microseismic waveform data, and energy abrupt changes exceeding 1×10⁻⁶ were extracted. 4 Precursor signals of rock strata fracturing at J / 10s;

[0017] Drift compensation for gas concentration time series data is performed using Kalman filtering, and sensor noise is corrected by predicting residuals using a long short-term memory network.

[0018] Preferably, the method for injecting the constraints of the rock mechanics equations is as follows:

[0019] Add a stress field energy conservation constraint term to the model loss function. This constraint term is defined as the physical consistency loss L. phy Its function is to constrain the physical phenomena predicted by the neural network through the law of conservation of energy in rock mechanics. The calculation formula is:

[0020]

[0021] Where, σ NN σ represents the stress field tensor predicted by the neural network. FEM This represents the stress field tensor calculated using finite element method (FEM) simulation. is the divergence operator used to quantify the energy conservation bias of the stress field; ||·|| represents the Euclidean norm used to calculate the difference between the predicted and simulated values; λ is the weighting coefficient with a value ranging from 0.05 to 0.2, determined through cross-validation.

[0022] Preferably, the cryptographic feature interactions of the federated learning framework include:

[0023] Homomorphic encryption is applied to the gradient parameters of the models generated locally in each mining area.

[0024] The central server aggregates and encrypts gradients and updates global model parameters.

[0025] Transmission of raw waveform data and gas concentration time series data is prohibited; only 256-dimensional encrypted feature vectors are shared.

[0026] Preferably, the process by which the digital twin simulation module generates disaster chain composite data includes:

[0027] Define the Mohr-Coulomb criterion for mining-induced stress field, Darcy's law for gas seepage field, and the extended finite element equation for fracture propagation.

[0028] Simulate the evolution path of a disaster chain, where rock bursts trigger the expansion of surrounding rock fissures, and these fissures facilitate gas outbursts and water level rises.

[0029] Label the synthetic data with disaster type labels, spatiotemporal location labels, and coupling strength labels.

[0030] Preferably, the formula for calculating the probability of the disaster chain propagation is:

[0031]

[0032] Among them, P i Let C represent the probability of occurrence of the i-th type of single disaster. ij represents the element in the i-th row and j-th column of the disaster chain coupling coefficient matrix, whose value is obtained by fitting historical disaster data, and Π represents the product of the joint probabilities of n disaster types.

[0033] Preferably, the preset threshold includes:

[0034] The triggering conditions for a yellow alert are dynamically adjusted through reinforcement learning, and are set as a single disaster probability greater than 60% or a disaster coupling probability greater than 30%.

[0035] The red alert is triggered when the disaster coupling probability is greater than 50% and the propagation speed exceeds 0.5m / s. The threshold is dynamically optimized by fitting historical data.

[0036] Preferably, the reinforcement learning mechanism includes:

[0037] Construct the reward function: R = a·accuracy - b·false alarm rate + c·response speed -1 ;

[0038] in:

[0039] Accuracy is defined as the proportion of disaster events correctly predicted by the model out of the total number of actual disaster events;

[0040] False alarm rate is defined as the proportion of non-disaster events that the model incorrectly warns of to the total number of actual non-disaster events;

[0041] Response speed is defined as the average time from when a sensor detects a precursor signal of a disaster to when the system generates an early warning command, in seconds;

[0042] a = 0.7, b = 0.3, and c = 0.2 are the accuracy weight, false alarm rate penalty coefficient, and response speed adjustment factor, respectively.

[0043] By dynamically evaluating the value of the reward function R, the reinforcement learning algorithm is driven to iteratively optimize the feature weights and warning thresholds of the multimodal deep learning model. Specific steps include:

[0044] Based on the comparison results between historical early warning records and actual disaster data, the accuracy, false alarm rate and response speed of the current model are calculated.

[0045] The near-end policy optimization algorithm is used to generate the policy gradient direction based on the reward function R, and then adjust the feature weights of each modality.

[0046] The weighting adjustment range for the time-series characteristics of gas concentration is:

[0047]

[0048] The weighting adjustment range for the microseismic waveform features is:

[0049]

[0050] Where, Δw gasThe adjustment amount for the time-series characteristic weights of gas concentration is calculated using the partial derivative of the reward function with respect to the weights. Let Δw be the partial derivative of the reward function with respect to the feature weights, representing the direction and magnitude of the impact of weight changes on the reward value. seismic The adjustment amount of the microseismic waveform feature weights is calculated by the partial derivative of the reward function with respect to the weights. η is the adaptive learning rate, which is dynamically adjusted according to the model convergence state. The initial value is 0.0003, which is used to control the parameter update step size.

[0051] Dynamically optimize early warning thresholds:

[0052] If the false alarm rate exceeds 15%, the threshold is lowered by ΔT = -β·(false alarm rate - 15%) to suppress false alarms;

[0053] If the false negative rate exceeds 5%, the threshold is increased by ΔT = +β·(false negative rate - 5%) to reduce false negatives;

[0054] Where β = 0.1 is the adjustment coefficient, representing the adjustment amount of the ΔT warning threshold;

[0055] The feature contribution heatmap is generated periodically and output through an interpretability verification interface to verify the rationality of the feature weight adjustment.

[0056] Preferably, the execution of the multi-level emergency control commands includes:

[0057] Send power-off and emergency ventilation start commands to the warning area;

[0058] By pushing disaster evacuation route planning maps dynamically generated based on the Dijkstra algorithm through augmented reality devices, the path update delay is less than 2 seconds;

[0059] A reinforcement instruction containing the grouting pressure threshold and material ratio parameters is issued to the grouting equipment.

[0060] Preferably, the optimization of the risk assessment and prevention strategy further includes:

[0061] Based on the disaster coupling probability output by the multimodal deep learning model, a heat map of risk level distribution is automatically generated.

[0062] Based on real-time microseismic energy data uploaded by edge computing devices, the priority of risk areas is dynamically adjusted;

[0063] By integrating unstructured geological exploration reports and sensor logs, text parsing, entity recognition, and data cleaning are performed to extract risk level keywords, spatial coordinate parameters, and disposal suggestion descriptions. After structured mapping, a safety situation data table containing risk level, spatial coordinates, and disposal suggestion codes is generated.

[0064] This invention provides an AI-based method for mine disaster prediction and risk assessment. It offers the following advantages:

[0065] 1. This invention constructs a multimodal deep learning model, integrating a one-dimensional convolutional neural network, a bidirectional long short-term memory network, and a graph attention network to extract microseismic waveforms, gas concentration time series, and spatial correlation features of geological structures, respectively. Rock mechanics equations are injected into the model training to constrain the model, effectively improving the feature extraction capability and fitting accuracy of disaster evolution laws for multi-source heterogeneous data under complex geological conditions. It solves the prediction bias problem caused by insufficient data fusion and lack of physical constraints in the existing technology, and significantly enhances the ability to identify precursor signals of rock strata rupture and the coupling process of disaster chains.

[0066] 2. This invention deploys a federated learning framework to homomorphically encrypt the gradient parameters of models trained locally in each mining area. The encrypted gradients are aggregated on a central server to update global model parameters. Only encrypted feature vectors are shared while the transmission of original data is prohibited, achieving a balance between cross-regional data sharing and privacy protection. At the same time, the digital twin simulation module is invoked to simulate the evolution path of disaster chains based on physical laws and generate labeled synthetic data, supplementing the scarce composite disaster samples in actual monitoring, breaking down data silos, and improving the model's generalization ability to complex geological conditions.

[0067] 3. This invention utilizes edge computing devices to detect energy mutations in real-time microseismic waveform data, quickly identifying precursory signals of rock strata rupture and reducing data transmission latency to within 50 milliseconds. Simultaneously, a cloud server calculates the probability of disaster chain propagation based on a multimodal model, forming a disaster prediction architecture that coordinates edge and cloud. When the probability exceeds a threshold, it triggers multi-level emergency control commands, including power shutdown, emergency ventilation adjustment, and grouting reinforcement parameter push. Furthermore, it dynamically plans evacuation routes using augmented reality devices, with path update latency of less than 2 seconds. This achieves millisecond-level closed-loop control from data acquisition to early warning execution, solving the problems of delayed early warning response and insufficient targeted emergency response in existing technologies, and significantly improving the timeliness and accuracy of disaster emergency handling.

[0068] 4. This invention constructs a reward function that includes accuracy, false alarm rate, and response speed. It uses an optimization algorithm to dynamically adjust the importance of different data features and early warning thresholds in the model. Combined with historical disaster data clustering analysis, it generates an intuitive risk heat map and dynamically updates the priority of risk areas based on real-time mining progress. At the same time, it integrates unstructured geological reports and sensor logs, transforming them into a structured safety situation data table containing risk level, spatial coordinates, and disposal suggestions. This upgrades risk assessment from static analysis to dynamic decision-making, enabling early warning thresholds and prevention and control strategies to be adaptively adjusted in real time according to geological changes and mining activities in the mining area. This solves the problems of fixed early warning strategies and lack of dynamic adjustment in risk assessment in existing technologies, significantly improving the real-time performance of risk assessment and the pertinence of prevention and control strategies, and providing more intelligent and accurate decision support for safe mine production. Attached Figure Description

[0069] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0070] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] Please see the appendix Figure 1 This invention provides an AI-based method for mine disaster prediction and risk assessment, comprising:

[0072] A multi-source sensor network is deployed, which includes microseismic monitoring sensors, gas concentration sensors, stress sensors and hydrological sensors, to collect microseismic waveform data, gas concentration time series data, stress data and hydrological parameter data of mining faces, roadways and geological structures.

[0073] Specifically, microseismic monitoring sensors, gas concentration sensors, stress sensors, and hydrological sensors are installed in the mining faces, roadways, and geological structural areas of the mine. The microseismic monitoring sensors use 1kHz high-frequency sampling to monitor rock fracture signals and achieve a positioning accuracy of ±5 meters. The gas concentration sensors track gas concentration changes in real time with a 0.5-second response time and a 10Hz update frequency. The stress sensors are placed at measuring points every 50 meters along the roadway roof to monitor surrounding rock stress with a resolution of 0.01MPa. The hydrological sensors use ultrasonic level gauges to detect water level changes in the goaf with an accuracy of ±2mm. Through this deployment, a multi-dimensional monitoring system covering key areas of the mine is formed, enabling precise location of microseismic events, second-level response to gas concentration anomalies, real-time early warning of stress changes, and dynamic tracking of hydrological parameters. This significantly improves the early identification capability of complex disasters and provides a high-quality data foundation for multimodal data analysis and intelligent early warning.

[0074] The microseismic waveform data, gas concentration time series data, stress data and hydrological parameter data are spatiotemporally aligned to generate a spatiotemporally aligned heterogeneous dataset, and encrypted feature interaction with external systems is performed through a federated learning framework.

[0075] Specifically, firstly, a dynamic time warping algorithm is used to match the time axis of 1kHz high-frequency waveform data from microseismic sensors, 10Hz time-series data from gas sensors, 1Hz low-frequency data from stress sensors, and 5-minute interval data of hydrological parameters, eliminating time offsets from multiple data sources. Then, energy accumulation features are extracted from the wavelet packet transform-denoised microseismic waveforms, and gas concentration drift errors are corrected using Kalman filtering. Furthermore, the sensor noise suppression rate is optimized by combining long short-term memory network prediction residuals. After cleaning and alignment, the heterogeneous dataset interacts with external mining area systems through a federated learning framework. The gradient parameters of locally trained models in each mining area are homomorphically encrypted and uploaded to the central server, aggregated to generate a global model, and then transmitted back to edge nodes, achieving cross-regional feature sharing while prohibiting the transmission of raw data. This step connects front-end multi-source data acquisition with back-end multi-modal model training. Spatiotemporal alignment improves data consistency and model input quality, while federated learning and encrypted interaction break down data silos and ensure privacy and security, ultimately supporting accurate calculation and dynamic early warning of disaster chain propagation probability.

[0076] A multimodal deep learning model is constructed based on the heterogeneous dataset. The multimodal deep learning model extracts microseismic waveform features through a one-dimensional convolutional neural network, extracts gas concentration temporal features through a bidirectional long short-term memory network, models the spatial correlation features of geological structures through a graph attention network, and injects rock mechanics equation constraints during the model training process.

[0077] Specifically, a one-dimensional convolutional neural network is used to process microseismic waveform data. Three convolutional layers with kernel sizes of 64, 32, and 16 are used to extract waveform features of rock strata fracture signals. A bidirectional long short-term memory network is used to analyze gas concentration time-series data, capturing the long-term dependence of concentration changes with a 128-element hidden layer. A graph attention network is used to model the relationships between geological structural nodes, generating correlation weights by combining the spatial locations of fault zones and goaf areas. During model training, rock mechanics equations are constrained and encoded as loss function terms, forcing the stress field predicted by the neural network to satisfy the law of energy conservation with the finite element simulation results. This step transforms multi-source heterogeneous data into a unified feature representation, enhancing the model's ability to fit the evolution of real disasters through physical constraints. This provides high-precision feature input for subsequent disaster chain synthesis data generation and supports cloud servers in accurately calculating the probability of disaster chain propagation and triggering dynamic early warnings.

[0078] The digital twin simulation module is invoked to generate disaster chain composite data, which includes evolution path label data of rock burst triggering surrounding rock fracture propagation, fracture conduction leading to abnormal gas outburst, and abrupt changes in hydrological parameters.

[0079] Specifically, based on the microseismic waveform features, gas concentration temporal patterns, and spatial correlation features of geological structures extracted by a multimodal deep learning model, and combined with the Mohr-Coulomb criterion for mining-induced stress fields, Darcy's law for gas seepage fields, and the extended finite element equation for fracture propagation, a high-fidelity digital twin of the mine is constructed. This simulates the chain evolution process of rockburst triggering fracture propagation in surrounding rock, abnormal gas outburst through fracture network conduction, and water level abrupt change caused by seepage field imbalance. Each simulated event is labeled with its disaster type, spatiotemporal coordinates, and coupling strength. This step supplements the deficiencies of actual disaster samples through a physics-driven simulation mechanism, enhancing the model's ability to learn the evolution patterns of complex disasters. The generated synthetic data is fused with real monitoring data and input into the multimodal model for joint training, significantly improving the generalization and reliability of subsequent disaster chain propagation probability calculations. This provides high-confidence data support for dynamic early warning threshold optimization and emergency decision-making.

[0080] Energy mutation detection is performed on real-time collected microseismic waveform data by edge computing devices. Events with energy mutations exceeding a preset threshold are identified as precursor signals of rock strata rupture. At the same time, the cloud server calculates the probability of disaster chain propagation based on the multimodal deep learning model.

[0081] Specifically, explosion-proof edge computing nodes are deployed in key underground areas. Wavelet packet transform algorithms are used to perform real-time noise reduction and energy accumulation calculations on 1kHz sampled microseismic waveforms. When an energy mutation exceeding 1×10⁻⁶ within a 10-second window is detected... 4Upon reaching the Joule threshold, the signal is immediately identified as a precursor to rock rupture, triggering a local alarm and uploading compressed and encrypted abnormal waveform segments to the cloud. Simultaneously, the cloud server receives microseismic features, gas concentration time-series patterns, and geological correlation weights from the multimodal deep learning model. Combining real-time data with historical disaster patterns, the probability of a compound disaster is calculated using the disaster chain propagation probability formula. This step seamlessly integrates front-end real-time detection with back-end model inference. Edge computing reduces data transmission latency to within 50 milliseconds, and the cloud model integrates multi-source features to dynamically extrapolate the disaster chain, providing millisecond-level response support for the precise triggering of subsequent multi-level emergency control commands, forming a closed-loop prevention and control chain from data acquisition to early warning execution.

[0082] When the probability of the disaster chain propagation exceeds a preset threshold, a multi-level emergency control command is triggered. The multi-level emergency control command includes a power shutdown command, an emergency ventilation parameter adjustment command, and a grouting reinforcement parameter push command.

[0083] Specifically, the cloud server monitors the disaster chain probability calculation results in real time. If the yellow alert conditions are met, a power shutdown command is sent to the underground equipment to cut off the power supply to the risk area. At the same time, the emergency ventilation system is adjusted to increase the wind speed to 3 meters per second to dilute the gas concentration. If the red alert conditions are met, an evacuation route based on real-time location is pushed through augmented reality devices, and reinforcement parameters of no less than 8 MPa and a water-cement ratio of 1:1 are sent to the grouting equipment. This step, based on the accurate calculation of the disaster chain probability at the front end, quickly blocks the disaster spread path through a graded response mechanism, providing real-time data support for subsequent risk assessment and prevention and control strategy optimization. At the same time, through closed-loop feedback, the execution effect data is sent back to the reinforcement learning module to drive the dynamic update of model parameters to improve the accuracy of future warnings.

[0084] Based on the comparison results between historical early warning records and actual disaster data, the feature weights and early warning thresholds of the multimodal deep learning model are dynamically adjusted through a reinforcement learning mechanism. Cluster analysis is performed on historical disaster data to generate risk heat maps, and the priority of risk areas is dynamically updated in conjunction with real-time mining progress to complete risk assessment and optimization of prevention and control strategies.

[0085] Specifically, based on the comparison between historical early warning records and actual disaster data, the model's accuracy, false alarm rate, and response speed are first calculated. A reward function incorporating these three indicators is constructed to evaluate the model's performance. An optimization algorithm-driven reinforcement learning mechanism is employed, dynamically adjusting the weights of features such as gas concentration time series and microseismic waveforms in the multimodal deep learning model based on reward function feedback. Simultaneously, the early warning threshold is automatically optimized based on exceeding false alarm or missed alarm rates, and the rationality of feature weight adjustments is periodically verified. Cluster analysis of historical disaster data generates risk heat maps. Combined with real-time microseismic energy data uploaded from edge devices and real-time mining progress, the priority of risk areas is dynamically updated, achieving an upgrade from static to dynamic risk assessment. This allows the early warning threshold to adapt to geological changes and mining activities in the mining area, improving the real-time nature and targeting of prevention and control strategies, and providing more intelligent decision support for safe mine production.

[0086] Spatiotemporal alignment processing includes:

[0087] A dynamic time warping algorithm was used to match the time axis of microseismic waveform data, gas concentration time series data, stress data and hydrological parameter data with different sampling frequencies.

[0088] Energy accumulation calculations were performed on the wavelet packet transform-denoised microseismic waveform data, and energy abrupt changes exceeding 1×10⁻⁶ were extracted. 4 Precursor signals of rock strata fracturing at J / 10s;

[0089] Drift compensation for gas concentration time series data is performed using Kalman filtering, and sensor noise is corrected by predicting residuals using a long short-term memory network.

[0090] Specifically, after deploying the multi-source sensor network and collecting microseismic waveforms, gas concentrations, stress, and hydrological data, a dynamic time warping algorithm is used to align data streams with different sampling frequencies to a unified time axis, eliminating time offsets in the 1kHz high-frequency microseismic signal, the 10Hz time-series gas data, the 1Hz low-frequency stress data, and the 5-minute interval hydrological data. Subsequently, energy accumulation calculations are performed on the wavelet packet transform-denoised microseismic waveforms, at a rate of 1×10^6 every 10 seconds. 4 Joule thresholding is used to screen for precursor signals of rock strata fracturing; simultaneously, Kalman filtering is used to correct drift errors in gas concentration sensors, and long short-term memory networks are used to predict temporal residuals to suppress electromagnetic interference noise. This step, as the core of data preprocessing, takes over the acquisition of raw data from the front end and provides high-quality input with spatiotemporal consistency and noise suppression for subsequent multimodal model training. Spatiotemporal alignment ensures the feature synchronization of multi-source data, energy mutation detection and drift compensation improve the extraction accuracy of disaster precursor signals, directly support encrypted feature interaction and model physical constraint injection under the federated learning framework, and ultimately enhance the reliability of disaster chain propagation probability calculation and the timeliness of early warning response.

[0091] The method for injecting constraints into the rock mechanics equations is as follows:

[0092] Add a stress field energy conservation constraint term to the model loss function. This constraint term is defined as the physical consistency loss L. phy Its function is to constrain the physical phenomena predicted by the neural network through the law of conservation of energy in rock mechanics. The calculation formula is:

[0093]

[0094] Where, σ NN σ represents the stress field tensor predicted by the neural network. FEM This represents the stress field tensor calculated using finite element method (FEM) simulation. is the divergence operator used to quantify the energy conservation bias of the stress field; ||·|| represents the Euclidean norm used to calculate the difference between the predicted and simulated values; λ is the weighting coefficient with a value ranging from 0.05 to 0.2, determined through cross-validation.

[0095] Specifically, during the training of the multimodal deep learning model, the constraints of the rock mechanics equations are encoded as loss function terms. The stress field distribution caused by mining is calculated through finite element simulation as a physical benchmark, forcing the stress field tensor predicted by the neural network to align with it in the energy conservation dimension. The weight coefficient λ is set to 0.05 to 0.2 through cross-validation to balance data-driven and physical laws. This step, based on the spatiotemporally aligned heterogeneous dataset, enhances the model's ability to fit the evolution of real disasters by fusing rock fracture mechanisms, reducing the risk of overfitting to sparse disaster samples. At the same time, it provides feature inputs that conform to mechanical principles for the subsequent digital twin simulation module, supporting the high-fidelity generation of disaster chain synthetic data, and ultimately improving the physical rationality of the cloud-based disaster chain propagation probability calculation and the credibility of early warning decisions.

[0096] The cryptographic feature interactions of the federated learning framework include:

[0097] Homomorphic encryption is applied to the gradient parameters of the models generated locally in each mining area.

[0098] The central server aggregates and encrypts gradients and updates global model parameters.

[0099] Transmission of raw waveform data and gas concentration time series data is prohibited; only 256-dimensional encrypted feature vectors are shared.

[0100] Specifically, after completing the spatiotemporal alignment and noise suppression of multi-source data, each mining area trains a multimodal deep learning model using heterogeneous datasets. The generated model gradient parameters are then converted into irreversible ciphertext using homomorphic encryption technology to ensure data privacy. After the encrypted gradients are uploaded to the central server, an aggregation algorithm performs weighted averaging and parameter fusion on the multi-source gradients to generate a globally optimized model, which is then distributed to each edge node for iterative updates. Simultaneously, the transmission of raw microseismic waveforms and gas concentration time-series data is strictly prohibited; only 256-dimensional encrypted feature vectors are allowed to participate in cross-mining area interactions. This step connects front-end data cleaning with back-end model training, breaking down data silos through a federated learning framework. While ensuring data sovereignty across mining areas, knowledge sharing is achieved. Encrypted gradient aggregation significantly improves the model's generalization ability to complex geological conditions, providing a highly robust model for cross-regional collaborative optimization for subsequent disaster chain propagation probability calculations. This also avoids the risk of sensitive information leakage, supporting the secure and reliable deployment of the mine disaster prediction system.

[0101] The process by which the digital twin simulation module generates composite disaster chain data includes:

[0102] Define the Mohr-Coulomb criterion for mining-induced stress field, Darcy's law for gas seepage field, and the extended finite element equation for fracture propagation.

[0103] Simulate the evolution path of a disaster chain, where rock bursts trigger the expansion of surrounding rock fissures, and these fissures facilitate gas outbursts and water level rises.

[0104] Label the synthetic data with disaster type labels, spatiotemporal location labels, and coupling strength labels.

[0105] Specifically, a rock mass fracture criterion is established based on the Mohr-Coulomb criterion of the mining-induced stress field. Darcy's law is used to simulate the gas diffusion path in the gas seepage field. The fracture propagation process is dynamically calculated through extended finite element equations. A digital twin model of the entire disaster evolution chain, including rockburst triggering surrounding rock fractures, fracture-induced abnormal gas outbursts, and water level rise, is constructed. During the simulation, the synthetic data is labeled with disaster type tags to distinguish between rockburst, gas outburst, and water inrush events. Spatiotemporal location tags record the disaster occurrence coordinates and timestamps, and coupling strength tags quantify the multi-hazard interaction levels. This step follows the physical constraint injection during the multimodal model training phase. High-fidelity simulation expands the sample of composite disasters scarce in actual monitoring. The generated labeled synthetic data is mixed with real data and input into the model for iterative optimization, directly supporting the accurate fitting of coupling coefficients in subsequent disaster chain propagation probability calculations. Simultaneously, it provides a multi-scenario extrapolation basis for the dynamic updating of risk assessment heatmaps, significantly improving the system's predictive coverage of rare disaster chain events and the adaptability of prevention and control strategies.

[0106] The formula for calculating the probability of a disaster chain propagation is:

[0107]

[0108] Among them, P i Let C represent the probability of occurrence of the i-th type of single disaster. ij represents the element in the i-th row and j-th column of the disaster chain coupling coefficient matrix, whose value is obtained by fitting historical disaster data, and Π represents the product of the joint probabilities of n disaster types.

[0109] Specifically, a disaster chain coupling coefficient matrix is ​​fitted using historical disaster data. Combined with real-time monitoring of microseismic energy, gas concentration, stress, and hydrological parameters, the probability of occurrence of each individual disaster is output through a multimodal deep learning model. This probability is then substituted into the disaster chain propagation probability formula for joint probability calculation. The total number of disaster types in the formula is dynamically adjusted according to the actual disaster type database of the mine. This step, based on front-end data acquisition and cleaning and multimodal feature extraction, quantifies and combines the high-confidence single-disaster probability output by the physical constraint model with the cross-disaster coupling effect. This directly supports the dynamic determination of subsequent multi-level early warning thresholds. The disaster chain cascading triggering mechanism is characterized by probability product, significantly improving the accuracy of predicting the evolution path of complex disasters. At the same time, it provides a quantitative basis for prioritizing emergency control instructions, achieving a leapfrog upgrade from single disaster monitoring to multi-disaster coupled risk assessment.

[0110] Preset thresholds include:

[0111] The triggering conditions for a yellow alert are dynamically adjusted through reinforcement learning, and are set as a single disaster probability greater than 60% or a disaster coupling probability greater than 30%.

[0112] The red alert is triggered when the disaster coupling probability is greater than 50% and the propagation speed exceeds 0.5m / s. The threshold is dynamically optimized by fitting historical data.

[0113] Specifically, the warning threshold setting is dynamically optimized through reinforcement learning algorithms. A reward function is trained based on historical disaster data and real-time monitoring results. Initially, a yellow warning is set as a single disaster probability greater than 60% or a coupling probability greater than 30%, and a red warning is set as a coupling probability greater than 50% and a propagation speed exceeding 0.5 meters per second. The threshold parameters are periodically updated using a near-end strategy optimization algorithm, and weight coefficients are adjusted using cross-validation to balance the false alarm rate and the missed alarm rate. This step connects the calculation of disaster chain propagation probability with multi-level emergency control. Dynamic thresholds adapt to changes in geological conditions and mining stages in different mining areas. Historical data fitting ensures that the warning conditions are consistent with the actual risk evolution pattern. The real-time optimization mechanism improves the sensitivity and reliability of the warning, directly triggering the precise execution of subsequent power shutdown, ventilation enhancement, and grouting reinforcement commands. Simultaneously, it provides a threshold benchmark for the dynamic updating of the risk assessment heatmap, achieving closed-loop management from probability quantification to graded response.

[0114] Reinforcement learning mechanisms include:

[0115] Construct the reward function: R = a·accuracy - b·false alarm rate + c·response speed -1 ;

[0116] in:

[0117] Accuracy is defined as the proportion of disaster events correctly predicted by the model out of the total number of actual disaster events;

[0118] False alarm rate is defined as the proportion of non-disaster events that the model incorrectly warns of to the total number of actual non-disaster events;

[0119] Response speed is defined as the average time from when a sensor detects a precursor signal of a disaster to when the system generates an early warning command, in seconds;

[0120] a = 0.7, b = 0.3, and c = 0.2 are the accuracy weight, false alarm rate penalty coefficient, and response speed adjustment factor, respectively.

[0121] By dynamically evaluating the value of the reward function R, the reinforcement learning algorithm is driven to iteratively optimize the feature weights and warning thresholds of the multimodal deep learning model. Specific steps include:

[0122] Based on the comparison results between historical early warning records and actual disaster data, the accuracy, false alarm rate and response speed of the current model are calculated.

[0123] The near-end policy optimization algorithm is used to generate the policy gradient direction based on the reward function R, and then adjust the feature weights of each modality.

[0124] The weighting adjustment range for the time-series characteristics of gas concentration is:

[0125]

[0126] The weighting adjustment range for the microseismic waveform features is:

[0127]

[0128] Where, Δw gas The adjustment amount for the time-series characteristic weights of gas concentration is calculated using the partial derivative of the reward function with respect to the weights. Let Δw be the partial derivative of the reward function with respect to the feature weights, representing the direction and magnitude of the impact of weight changes on the reward value. seismic The adjustment amount of the microseismic waveform feature weights is calculated by the partial derivative of the reward function with respect to the weights. η is the adaptive learning rate, which is dynamically adjusted according to the model convergence state. The initial value is 0.0003, which is used to control the parameter update step size.

[0129] Dynamically optimize early warning thresholds:

[0130] If the false alarm rate exceeds 15%, the threshold is lowered by ΔT = -β·(false alarm rate - 15%) to suppress false alarms;

[0131] If the false negative rate exceeds 5%, the threshold is increased by ΔT = +β·(false negative rate - 5%) to reduce false negatives;

[0132] Where β = 0.1 is the adjustment coefficient, representing the adjustment amount of the ΔT warning threshold;

[0133] The feature contribution heatmap is generated periodically and output through an interpretability verification interface to verify the rationality of the feature weight adjustment.

[0134] Specifically, the reinforcement learning mechanism dynamically optimizes the parameters of the multimodal deep learning model by constructing a reward function. Accuracy and false alarm rate are defined as the ratio of correct to incorrect warnings, respectively, and response speed is the delay time from sensor detection to warning generation. A proximal strategy optimization algorithm is employed, adjusting the weights of gas concentration temporal features and microseismic waveform features according to the gradient direction of the reward function, with an adaptive learning rate controlling the parameter update step size. Simultaneously, the warning threshold is dynamically adjusted: if the false alarm rate exceeds 15%, the threshold is lowered to suppress false alarms; if the missed alarm rate exceeds 5%, the threshold is raised to reduce missed alarms, with adjustment coefficients controlling the adjustment magnitude. Periodically generated feature contribution heatmaps verify the rationality of the weights. This mechanism, by balancing accuracy, false alarm rate, and response speed, enables the model to adapt to dynamic changes in the mining area, increasing warning accuracy to 92%, reducing the false alarm rate to below 12%, and shortening the response speed to within 3 seconds, significantly enhancing the real-time performance and reliability of disaster prediction and providing support for accurate emergency decision-making.

[0135] The execution of multi-level emergency control commands includes:

[0136] Send power-off and emergency ventilation start commands to the warning area;

[0137] By pushing disaster evacuation route planning maps dynamically generated based on the Dijkstra algorithm through augmented reality devices, the path update delay is less than 2 seconds;

[0138] A reinforcement instruction containing the grouting pressure threshold and material ratio parameters is issued to the grouting equipment.

[0139] Specifically, when the cloud server determines that the probability of a disaster chain propagation exceeds a preset threshold, it immediately sends a power-off command to the warning area to cut off the power supply network to the dangerous area. Simultaneously, the emergency ventilation system is activated to adjust the wind speed to 3 meters per second to accelerate gas dilution. The location of personnel underground is obtained in real time through the positioning module built into the augmented reality device. Based on the Dijkstra algorithm, the shortest evacuation route is dynamically planned and a 3D navigation map is generated, ensuring that the route update delay is less than 2 seconds. At the same time, commands containing an 8 MPa grouting pressure threshold and a 1:1 water-cement ratio are sent to the grouting equipment, driving the grouting pump to perform surrounding rock reinforcement according to the optimized parameters. This execution mechanism eliminates the risk of electrical sparks through rapid power cut-off, improves personnel evacuation efficiency by more than 30% through dynamic evacuation routes, and increases the surrounding rock reinforcement strength by 25% through precise grouting parameters. The multi-level command coordination forms a closed-loop prevention and control chain from early warning to response, effectively blocking the spread of the disaster chain.

[0140] Risk assessment and prevention strategy optimization also includes:

[0141] Based on the disaster coupling probability output by the multimodal deep learning model, a heat map of risk level distribution is automatically generated.

[0142] Based on real-time microseismic energy data uploaded by edge computing devices, the priority of risk areas is dynamically adjusted;

[0143] By integrating unstructured geological exploration reports and sensor logs, text parsing, entity recognition, and data cleaning are performed to extract risk level keywords, spatial coordinate parameters, and disposal suggestion descriptions. After structured mapping, a safety situation data table containing risk level, spatial coordinates, and disposal suggestion codes is generated.

[0144] Specifically, disaster coupling probability data output by a multimodal deep learning model is used to classify the risk levels of the mine's three-dimensional spatial coordinates using the K-means clustering algorithm, generating a dynamic heat map. Heat values ​​are mapped to the mining area using a four-color gradient (red, orange, yellow, blue). Edge computing devices upload microseismic energy monitoring data in real time, and the cloud server refreshes the risk area priority ranking every 5 minutes. Combined with the mining progress model, the risk diffusion path is predicted for the next 8 hours. Unstructured geological exploration reports undergo entity recognition using natural language processing technology, extracting fault zone coordinates, rock permeability, and historical accident keywords. After fusing with real-time parameters from sensor logs, data cleaning and structured mapping generate a safety situation data table containing risk levels, spatial coordinates, and handling codes. This method increases the mine's composite risk identification coverage to 95%, reduces the risk area priority update delay to less than 5 seconds, improves the efficiency of prevention and control strategy formulation supported by the structured data table by 40%, achieves a spatial resolution of 5 meters for the risk heat map, and increases the accuracy of abnormal area location to over 92%, effectively guiding underground resource scheduling and dynamic optimization of emergency plans.

[0145] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based method for mine disaster prediction and risk assessment, characterized in that, include: A multi-source sensor network is deployed, which includes microseismic monitoring sensors, gas concentration sensors, stress sensors and hydrological sensors, to collect microseismic waveform data, gas concentration time series data, stress data and hydrological parameter data of mining faces, roadways and geological structures. The microseismic waveform data, gas concentration time series data, stress data and hydrological parameter data are spatiotemporally aligned to generate a spatiotemporally aligned heterogeneous dataset, and encrypted feature interaction with external systems is performed through a federated learning framework. A multimodal deep learning model is constructed based on the heterogeneous dataset. The multimodal deep learning model extracts microseismic waveform features through a one-dimensional convolutional neural network, extracts gas concentration temporal features through a bidirectional long short-term memory network, models the spatial correlation features of geological structures through a graph attention network, and injects rock mechanics equation constraints during the model training process. The digital twin simulation module is invoked to generate disaster chain composite data, which includes evolution path label data of rock burst triggering surrounding rock fracture propagation, fracture conduction leading to abnormal gas outburst, and abrupt changes in hydrological parameters. Energy mutation detection is performed on real-time collected microseismic waveform data by edge computing devices. Events with energy mutations exceeding a preset threshold are identified as precursor signals of rock strata rupture. At the same time, the cloud server calculates the probability of disaster chain propagation based on the multimodal deep learning model. When the probability of the disaster chain propagation exceeds a preset threshold, a multi-level emergency control command is triggered. The multi-level emergency control command includes a power shutdown command, an emergency ventilation parameter adjustment command, and a grouting reinforcement parameter push command. Based on the comparison results between historical early warning records and actual disaster data, the feature weights and early warning thresholds of the multimodal deep learning model are dynamically adjusted through a reinforcement learning mechanism. Cluster analysis is performed on historical disaster data to generate risk heat maps, and the priority of risk areas is dynamically updated in conjunction with real-time mining progress to complete risk assessment and optimization of prevention and control strategies.

2. The method for mine disaster prediction and risk assessment based on AI according to claim 1, characterized in that, The spatiotemporal alignment process includes: A dynamic time warping algorithm was used to match the time axis of microseismic waveform data, gas concentration time series data, stress data and hydrological parameter data with different sampling frequencies. Energy accumulation calculations were performed on the wavelet packet transform-denoised microseismic waveform data, and energy abrupt changes exceeding 1×10⁻⁶ were extracted. 4 Precursor signals of rock strata fracturing at J / 10s; Drift compensation for gas concentration time series data is performed using Kalman filtering, and sensor noise is corrected by predicting residuals using a long short-term memory network.

3. The method for mine disaster prediction and risk assessment based on AI according to claim 1, characterized in that, The method for injecting the constraints of the rock mechanics equations is as follows: Add a stress field energy conservation constraint term to the model loss function. This constraint term is defined as the physical consistency loss L. phy Its function is to constrain the physical phenomena predicted by the neural network through the law of conservation of energy in rock mechanics. The calculation formula is: Where, σ NN σ represents the stress field tensor predicted by the neural network. FEM This represents the stress field tensor calculated using finite element method (FEM) simulation. is the divergence operator used to quantify the energy conservation bias of the stress field; ||·|| represents the Euclidean norm used to calculate the difference between the predicted and simulated values; λ is the weighting coefficient with a value ranging from 0.05 to 0.2, determined through cross-validation.

4. The method for mine disaster prediction and risk assessment based on AI according to claim 1, characterized in that, The cryptographic feature interactions of the federated learning framework include: Homomorphic encryption is applied to the gradient parameters of the models generated locally in each mining area. The central server aggregates and encrypts gradients and updates global model parameters. Transmission of raw waveform data and gas concentration time series data is prohibited; only 256-dimensional encrypted feature vectors are shared.

5. The method for mine disaster prediction and risk assessment based on AI according to claim 1, characterized in that, The process by which the digital twin simulation module generates composite disaster chain data includes: Define the Mohr-Coulomb criterion for mining-induced stress field, Darcy's law for gas seepage field, and the extended finite element equation for fracture propagation. Simulate the evolution path of a disaster chain, where rock bursts trigger the expansion of surrounding rock fissures, and these fissures facilitate gas outbursts and water level rises. Label the synthetic data with disaster type labels, spatiotemporal location labels, and coupling strength labels.

6. The method for mine disaster prediction and risk assessment based on AI according to claim 1, characterized in that, The formula for calculating the probability of the disaster chain propagation is: Among them, P i Let C represent the probability of occurrence of the i-th type of single disaster. ij represents the element in the i-th row and j-th column of the disaster chain coupling coefficient matrix, whose value is obtained by fitting historical disaster data, and Π represents the product of the joint probabilities of n disaster types.

7. The method for mine disaster prediction and risk assessment based on AI according to claim 1, characterized in that, The preset threshold includes: The triggering conditions for a yellow alert are dynamically adjusted through reinforcement learning, and are set as a single disaster probability greater than 60% or a disaster coupling probability greater than 30%. The red alert is triggered when the disaster coupling probability is greater than 50% and the propagation speed exceeds 0.5m / s. The threshold is dynamically optimized by fitting historical data.

8. The method for mine disaster prediction and risk assessment based on AI according to claim 1, characterized in that, The reinforcement learning mechanism includes: Construct the reward function: R = a·accuracy - b·false alarm rate + c·response speed -1 ; in: Accuracy is defined as the proportion of disaster events correctly predicted by the model out of the total number of actual disaster events; False alarm rate is defined as the proportion of non-disaster events that the model incorrectly warns of to the total number of actual non-disaster events; Response speed is defined as the average time from when a sensor detects a precursor signal of a disaster to when the system generates an early warning command, in seconds; a = 0.7, b = 0.3, and c = 0.2 are the accuracy weight, false alarm rate penalty coefficient, and response speed adjustment factor, respectively. By dynamically evaluating the value of the reward function R, the reinforcement learning algorithm is driven to iteratively optimize the feature weights and warning thresholds of the multimodal deep learning model. Specific steps include: Based on the comparison results between historical early warning records and actual disaster data, the accuracy, false alarm rate and response speed of the current model are calculated. The near-end policy optimization algorithm is used to generate the policy gradient direction based on the reward function R, and then adjust the feature weights of each modality. The weighting adjustment range for the time-series characteristics of gas concentration is: The weighting adjustment range for the microseismic waveform features is: Where, Δw gas The adjustment amount for the time-series characteristic weights of gas concentration is calculated using the partial derivative of the reward function with respect to the weights. Let Δw be the partial derivative of the reward function with respect to the feature weights, representing the direction and magnitude of the impact of weight changes on the reward value. seismic The adjustment amount of the microseismic waveform feature weights is calculated by the partial derivative of the reward function with respect to the weights. η is the adaptive learning rate, which is dynamically adjusted according to the model convergence state. The initial value is 0.0003, which is used to control the parameter update step size. Dynamically optimize early warning thresholds: If the false alarm rate exceeds 15%, the threshold is lowered by ΔT = -β·(false alarm rate - 15%) to suppress false alarms; If the false negative rate exceeds 5%, the threshold is increased by ΔT = +β·(false negative rate - 5%) to reduce false negatives; Where β = 0.1 is the adjustment coefficient, representing the adjustment amount of the ΔT warning threshold; The feature contribution heatmap is generated periodically and output through an interpretability verification interface to verify the rationality of the feature weight adjustment.

9. The method for mine disaster prediction and risk assessment based on AI according to claim 1, characterized in that, The execution of the multi-level emergency control commands includes: Send power-off and emergency ventilation start commands to the warning area; By pushing disaster evacuation route planning maps dynamically generated based on the Dijkstra algorithm through augmented reality devices, the path update delay is less than 2 seconds; A reinforcement instruction containing the grouting pressure threshold and material ratio parameters is issued to the grouting equipment.

10. The AI-based method for mine disaster prediction and risk assessment according to claim 1, characterized in that, The optimization of risk assessment and prevention strategies also includes: Based on the disaster coupling probability output by the multimodal deep learning model, a heat map of risk level distribution is automatically generated. Based on real-time microseismic energy data uploaded by edge computing devices, the priority of risk areas is dynamically adjusted; unstructured geological exploration reports and sensor logs are integrated, and text parsing, entity recognition and data cleaning are performed to extract risk level keywords, spatial coordinate parameters and disposal suggestion descriptions. After structured mapping, a safety situation data table containing risk level, spatial coordinates and disposal suggestion codes is generated.

Images