Intelligent mine pressure monitoring method and device based on multi-modal data fusion and electronic equipment

Abstract

The application relates to the technical field of coal mine safety monitoring, and discloses a multi-modal data fusion digital mine pressure monitoring method, device and electronic equipment. The method first drives distributed optical fibers, microseismic detectors, cameras and other heterogeneous sensors through a data acquisition module to synchronously acquire overburden movement and support response data of a fully mechanized working face. Subsequently, the data processing module completes cleaning and feature extraction, and inputs the data into an intelligent analysis module with an embedded attention mechanism and a long short-term memory network to deeply mine the space-time evolution law of mine pressure and output real-time pressure probability. The system determines the risk level according to the probability interval and executes hierarchical linkage: low risk maintains normal monitoring, medium risk prompts strengthened patrol, and high risk directly triggers an audible and light alarm and automatically interlocks to cut off the equipment power to forcibly evacuate people. The application effectively overcomes the limitations of single-source monitoring, significantly improves the accuracy of roof disaster early warning and the timeliness of emergency response through multi-modal deep fusion and automatic interlocking control.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety monitoring technology, specifically to a digital intelligent mine pressure monitoring method, device, and electronic equipment based on multimodal data fusion. Background Technology

[0002] Periodic roof pressure in fully mechanized mining faces is one of the major hazards restricting safe and efficient coal mine production. Accurately identifying signs of roof pressure and taking timely preventive measures is crucial to preventing roof accidents. Currently, monitoring of roof pressure mainly relies on hydraulic support resistance monitoring or a single microseismic monitoring system. However, relying solely on support resistance data can only reflect the passive interaction force between the roof and the support, failing to deeply perceive the precursors of fracture evolution within the overburden. Furthermore, due to the harsh underground environment, pressure sensors are easily affected by the operating conditions of the hydraulic system itself, resulting in significant lag and limitations in the monitoring data. Meanwhile, although microseismic monitoring can capture the energy signals released by rock fractures, it is highly sensitive to mechanical vibrations and noise from on-site coal cutting machines and scraper conveyor operation. Without video or other intuitive data for verification, it is easily confused with valid signals and environmental noise, leading to frequent false alarms. Furthermore, existing monitoring systems mostly employ independent data processing modes, failing to effectively integrate multi-source heterogeneous data such as stress, microseismic data, and video. Their early warning logic is often based on simple static threshold judgments, lacking in-depth learning and analysis of the spatiotemporal evolution trends of mine pressure, making it difficult to distinguish the intensity levels of incoming pressure. When dealing with sudden, strong incoming pressure risks, existing alarm mechanisms often stop at audible and visual alerts, lacking automatic interlocking control functions with fully mechanized mining equipment. This reliance on manual decision-making and operation often results in delays in emergency situations, failing to meet the timeliness requirements for on-site emergency evacuation. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a multimodal data fusion-based intelligent mining pressure monitoring method, device, and electronic equipment. This solves the problems of low early warning accuracy and high false alarm rate caused by reliance on a single monitoring source in existing mining pressure monitoring methods, as well as the lack of an automatic interlocking mechanism with fully mechanized mining equipment, which leads to delayed emergency response.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The first invention provides a multimodal data fusion-based intelligent mine pressure monitoring method, comprising the following steps: S1. Construct a multi-source heterogeneous sensing network. Based on the geological conditions of the mining face and roadway, deploy various monitoring devices including distributed fiber optic sensors, microseismic detectors, displacement sensors, stress sensors, and cameras. The cameras are installed on hydraulic supports to form a full-space monitoring network covering the working face and surrounding rock mass. S2. Multimodal data acquisition and transmission: Real-time acquisition of internal stress, vibration signals, displacement data and coal wall fracture images of the rock mass through the sensing network, and transmission of the acquired data to the cloud platform through the edge computing gateway; S3. Data preprocessing: Denoising is performed on data of different modalities, and abnormal or missing data is repaired to obtain standardized time series data. S4. Multidimensional spatiotemporal feature extraction: Extracting temporal features that characterize the precursors of roof pressure from the preprocessed data, including stress statistical features, displacement motion features, microseismic energy features, and crack image features. S5. Construct an intelligent early warning model. Based on historical data, build a deep learning model that includes an attention mechanism and a long short-term memory network. Use the attention mechanism to dynamically weight and fuse multimodal features. Use the long short-term memory network to capture the temporal evolution pattern and output the probability and risk level of the top plate. S6. Tiered early warning and linkage response: Based on the risk level output by the model, the corresponding early warning signal and on-site handling measures are automatically triggered. S7. Feedback optimization: Compare the actual mining pressure events with the early warning results, and use the comparison results to adjust the parameters of the intelligent early warning model.

[0005] Preferably, the deployment of the sensing network in step S1 specifically includes: Distributed fiber optic sensors are embedded in the roadway roof and coal face to monitor the stress distribution inside the rock mass. Microseismic detectors were installed on the roof and sidewalls of the two roadways in the working face to collect vibration information of rock mass fracture. Displacement sensors are placed on the roof of the roadway and in both roadways to monitor the displacement of the coal and rock mass and the deformation of the surrounding rock. Stress sensors are installed inside the hydraulic support column and in the tunnel borehole to collect support pressure and bearing pressure. The camera is installed below the top beam of the hydraulic support, with the lens facing the coal face, to monitor the development of cracks on the coal face surface and the contact status between the support and the roof.

[0006] Preferably, the specific method of data preprocessing in step S3 is as follows: For stress and displacement data, a state-space model is established, and the Kalman filter algorithm is used for noise reduction. For microseismic signals, wavelet threshold denoising method is used to remove low-frequency interference components through wavelet decomposition; For the image data, Gaussian filtering is used to remove noise, and histogram equalization is used to adjust the image contrast. The surface areas of the coal wall and roof are extracted through image segmentation. For suspicious or missing data marked in the initial verification, if the deviation is within the preset range, interpolation is used to repair it; if the deviation exceeds the preset range or the data is missing, a time series prediction model is used to supplement the prediction based on historical data trends.

[0007] Preferably, the feature extraction in step S4 specifically includes: Stress-related characteristics: Calculate the rate of change of stress per unit time, the amplitude of stress fluctuation, and the mean and standard deviation of stress; Displacement-related characteristics: Calculate displacement acceleration, cumulative displacement, and displacement change frequency; Microseismic characteristics: Calculate the number of microseismic events per unit time, total microseismic energy, average energy, and the proportion of high-frequency microseismic events; Image-related features: Cracks on the coal face and roof surface are extracted using image recognition algorithms, and the crack density change rate, average crack length, and average crack width are calculated.

[0008] Preferably, the construction and operation of the intelligent early warning model in step S5 includes: Feature fusion: The extracted multidimensional feature vector is input into the attention mechanism module, the importance weight of each feature to the prediction of the top plate pressure is calculated, and the fused feature vector is generated according to the weight. Temporal prediction: The fused feature vectors are input into a long short-term memory network according to the time series, and the evolution law of mining pressure is learned by using a gating mechanism; Risk Classification: The model outputs the probability value of the roof slab pressing down within a preset time period in the future, and determines the risk level based on the interval to which the probability value belongs.

[0009] Preferably, the model training process in step S5 includes: A dataset was constructed, and historical monitoring data was divided into non-pressure samples and periodic pressure samples. The periodic pressure samples were further subdivided into weak pressure, medium pressure and strong pressure samples. The model is trained using the pre-defined training set, the loss is calculated using the cross-entropy loss function, and the model parameters are updated using the backpropagation algorithm. Regularization techniques are introduced to randomly discard some network nodes to suppress overfitting, and the model performance is evaluated using a validation set. Training stops when the performance metrics meet the preset requirements.

[0010] Preferably, the specific threshold for risk classification is set as follows: When the probability of pressure is less than or equal to 0.3, it is judged as a low-risk level; When the probability of pressure is greater than 0.3 and less than or equal to 0.7, it is judged as a medium-risk level; When the probability of pressure is greater than 0.7, it is judged as a high-risk level.

[0011] Preferably, the graded early warning and linkage response in step S6 includes: When the risk level is determined to be low, the system will display a normal message and will not initiate active alarms. When the risk level is determined to be medium, a voice broadcast and pop-up reminder will be issued, the sensor sampling frequency will be increased, and personnel will be notified to strengthen on-site inspections. When a high-risk level is determined, an audible and visual alarm and a telephone call will be activated. The system will automatically interlock to prevent equipment in the alarm area from performing unnecessary actions, and will forcibly stop operations and evacuate personnel.

[0012] The second invention provides a multimodal data fusion-based intelligent mine pressure monitoring device, comprising: The data acquisition module is used to drive distributed fiber optic sensors, micro-vibration detectors, displacement sensors, stress sensors, and cameras to acquire data. The data processing module is configured to perform data preprocessing and feature extraction steps. The intelligent analysis module has a built-in early warning model based on attention mechanism and long short-term memory network, which is used to perform multimodal feature fusion and pressure probability prediction. The early warning execution module is used to trigger corresponding alarm devices and control commands based on the risk level output by the intelligent analysis module.

[0013] The second invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor.

[0014] Working Principle: During operation, the data acquisition module serves as the underlying hardware interface, driving the distributed fiber optic sensors, microseismic detectors, displacement sensors, stress sensors, and cameras located on the longwall mining face to operate synchronously. This module acquires vibration signals from the overburden, the stress state of the supports, and on-site video images according to a unified clock reference, converting multi-source heterogeneous physical signals into digital signals and resolving the time synchronization problem between different sensors. Subsequently, the data processing module receives the raw data, uses filtering algorithms to remove environmental noise interference from underground mining equipment, and transforms unstructured data into machine-readable feature vectors, completing data cleaning and feature extraction. The preprocessed feature data is input to the intelligent analysis module. This module runs a deep learning model based on attention mechanisms and long short-term memory networks, automatically fusing and analyzing multi-dimensional features. By mining the evolutionary patterns of the data over time, the model identifies precursory features of roof pressure and calculates the probability of pressure under the current working conditions in real time. Finally, the early warning execution module receives the probability value in real time and compares it with the preset risk threshold to determine the current risk level and execute graded linkage control: when the risk is low, the device only displays and records the data on the screen; when the risk is medium, the device triggers voice prompts and pop-up warnings to remind personnel to pay attention and strengthen support, while increasing the sampling frequency of the sensors for key monitoring; when the risk is high, the device immediately activates the high-intensity audible and visual alarm on site and outputs automatic control commands to the fully mechanized mining equipment control center, forcibly triggering the interlocking mechanism to cut off the power source of key equipment such as the coal mining machine, thereby achieving proactive risk avoidance by shutting down equipment and evacuating personnel before strong pressure occurs.

[0015] This invention provides a multimodal data fusion-based intelligent mining pressure monitoring method, device, and electronic equipment. It offers the following advantages: 1. This invention overcomes the shortcomings of single-source monitoring information by integrating multimodal data such as distributed optical fiber, microseismic data, and video surveillance. By combining attention mechanisms and long short-term memory networks, the system can automatically filter key precursor features and mine the evolutionary patterns in time series, effectively distinguishing between real mine pressure manifestations and environmental noise interference, significantly improving the accuracy of roof pressure prediction, and solving the problems of high false alarm rate and high risk of missed alarm caused by the single feature of traditional monitoring methods.

[0016] 2. This invention employs a risk grading strategy based on probability intervals, setting clear thresholds for low, medium, and high risks. By distinguishing different risk levels, the system can maintain normal operations during low-risk periods, prompt for enhanced inspections and support adjustments during medium-risk periods, and trigger mandatory measures only during high-risk periods. This refined grading management avoids frequent shutdowns caused by minor disturbances in the rock mass, maximizing both production continuity and efficiency of the fully mechanized mining face while strictly ensuring safety.

[0017] 3. This invention establishes an automatic interlocking control mechanism between monitoring and early warning results and on-site equipment. When a high-risk level is determined, the system can directly trigger an audible and visual alarm and forcibly cut off the power source of key equipment such as coal mining machines, without the need for manual reporting and confirmation at each level. This eliminates the risk of delayed human decision-making or poor communication in emergency situations, ensuring that equipment shutdown and personnel evacuation are completed within the critical window before strong pressure occurs, providing reliable active safety protection for underground operations. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a sensor arrangement diagram of the present invention.

[0019] Among them, 1. Distributed fiber optic sensor; 2. Micro-vibration detector; 3. Displacement sensor; 4. Stress sensor; 5. Camera; 6. Hydraulic support. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below 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.

[0021] Example: Please see the appendix Figure 1 This invention provides a multimodal data fusion-based intelligent mining pressure monitoring method, comprising the following steps: S1. Construct a multi-source heterogeneous sensing network. Based on the geological conditions of the mining face and roadway, deploy a variety of monitoring devices, including distributed fiber optic sensors 1, microseismic detectors 2, displacement sensors 3, stress sensors 4, and cameras 5. Cameras 5 are installed on hydraulic supports 6 to form a full-space monitoring network covering the working face and surrounding rock mass.

[0022] The deployment of the sensing network in step S1 specifically includes: Distributed fiber optic sensor 1 is embedded in the direction of the roadway roof and the coal wall to monitor the stress distribution inside the rock mass; Microseismic detector 2 is installed on the roof and sidewalls of the two roadways of the working face to collect vibration information of rock mass fracture. Displacement sensor 3 is placed on the roof of the roadway and in both roadways to monitor the displacement of the coal and rock mass and the deformation of the surrounding rock. Stress sensor 4 is installed inside the hydraulic support 6 column and in the tunnel borehole to collect support pressure and bearing pressure. Camera 5 is installed below the top beam of hydraulic support 6, with the lens facing the coal wall, to monitor the development of cracks on the coal wall surface and the contact status between the support and the roof.

[0023] This deployment scheme utilizes photosensitive elements to continuously sense minute deformations in the rock mass, capturing changes in the internal stress field, effectively eliminating blind spots in single-point monitoring and improving spatial resolution. Simultaneously, by capturing elastic wave signals, it locates microseismic sources, providing dynamic support for assessing rock damage and fracture trends. In terms of stability monitoring, real-time quantification of settlement and delamination helps to promptly detect abnormal deformations to prevent collapse and ensure operational safety. Furthermore, stress monitoring of the support structure provides direct mechanical basis for determining pressure intensity and optimizing support parameters. In addition, visual monitoring combined with image recognition technology intuitively records crack evolution, filling gaps in texture feature data and significantly improving the ability to identify risks such as end-face leakage.

[0024] S2. Multimodal data acquisition and transmission: Real-time acquisition of internal rock stress, vibration signals, displacement data, and coal wall fracture images via a sensing network, and transmission of the acquired data to the cloud platform via an edge computing gateway.

[0025] S3. Data preprocessing involves denoising the data from different modalities and repairing any abnormal or missing data to obtain standardized time-series data. The specific method of data preprocessing in step S3 is as follows: Numerical time series data noise reduction: To address the random noise present in the stress and displacement data, a state-space model is established, and the Kalman filter algorithm is used for noise reduction. Specifically, the state equation is defined as follows: The observation equation is .in, for The system state at any given moment. For the observed values, Here is the state transition matrix. For the observation matrix, and These are process noise and observation noise, respectively. The recursive filtering process includes prediction and updating: Prior state estimation: ; Prior error covariance: ; Kalman gain: ; Posterior state estimation: ; Posterior error covariance: ; By iterating through the above steps, the denoised optimal state estimation sequence is obtained. .

[0026] To address background noise such as equipment vibration mixed in microseismic signals, wavelet threshold denoising method is used.

[0027] First, the noisy signal conduct Layer wavelet decomposition: ; in, For the first Layer approximation coefficient, For the first Layer detail factor.

[0028] Then, the detail coefficients for each layer were analyzed. Apply a soft threshold function to process: ; in, The threshold is typically set according to the Donoho-Johnstone criterion. Finally, the processed coefficients are subjected to inverse wavelet transform to reconstruct the denoised microseismic signal.

[0029] Image data preprocessing: For image data, a two-dimensional zero-mean Gaussian function is first used. Used as a convolution kernel for Gaussian filtering and noise reduction: ; For images Smooth Gaussian noise using convolution operations: .

[0030] Then, histogram equalization is used to enhance contrast, and the cumulative distribution function (CDF) is calculated: ; And perform grayscale mapping: Finally, image segmentation techniques were used to extract the surface areas of the coal face and roof.

[0031] Repairing anomalies and missing data: For suspicious or missing data identified in the initial verification: If the data deviation is within the preset range, linear interpolation will be used for repair. ; If the data deviation exceeds the preset range or data is missing, the ARIMA model is used to supplement the time series forecast: ; in, As a lag operator, this model is used to predict data trends during missing periods, thus maintaining sequence continuity.

[0032] This preprocessing workflow first utilizes the Kalman filter algorithm to eliminate random errors and environmental noise, significantly improving the accuracy of the data in reflecting the true stress state of the rock mass. Simultaneously, wavelet thresholding denoising is employed to filter out low-frequency interference caused by downhole mechanical vibrations, effectively preserving high-frequency fracture characteristics and improving the signal-to-noise ratio. Regarding visual data, smoothing particle noise and expanding the grayscale range resolves the blurring issues caused by dust and uneven lighting, successfully highlighting fracture texture. Finally, a data restoration strategy ensures the integrity of the time series and the continuity of feature extraction, effectively mitigating the risk of model misjudgment due to data interruption.

[0033] S4. Multidimensional Spatiotemporal Feature Extraction: Extracting temporal features that characterize the precursors of roof pressure from the preprocessed data.

[0034] The feature extraction in step S4 specifically includes: Stress-related characteristics: Calculate the rate of change of stress This reflects the rate of stress accumulation in the rock mass. ; in This represents the rate of change of displacement.

[0035] Calculate the stress fluctuation amplitude That is, the difference between the peak value and the trough value per unit time: ; Simultaneously, the average stress per unit time is statistically analyzed. with standard deviation .

[0036] Displacement-related characteristics: Calculate displacement acceleration This reflects the accelerating trend of roof subsidence: ; in This represents the rate of change of displacement.

[0037] Calculate the cumulative displacement : ; And the frequency of displacement change is statistically analyzed.

[0038] Microseismic characteristics: Count of microseismic events per unit time .

[0039] Calculate the total energy of microseismic events ,in For the first The energy of each event was calculated, along with the proportion of high-frequency microseismic events. .

[0040] Image-related features: The cracks on the coal face and roof surface were extracted using image recognition algorithms, and the crack density change rate was calculated. : ; in To increase the number of fractures, The monitored area is used as a reference. The average fracture length is also calculated. and average crack width .

[0041] The extracted stress-related features aim to quantify the dynamic evolution of loads within the rock mass. By reflecting the degree of stress concentration and nonlinear abrupt change trends, they construct mechanical precursor indicators for identifying dynamic disasters. Combined with displacement-related features to describe the spatial motion state of the surrounding rock, they can accurately capture the critical point of the roof transitioning from a steady state to an unstable state, thereby effectively predicting the risk of large deformations. Simultaneously, microseismic-related features characterize the level of damage accumulation and energy release deep within the rock mass, providing geophysical evidence for assessing fracture trends by inverting fracture propagation activity. Image-related features intuitively quantify the degree of shallow surface damage, significantly improving the identification accuracy of disasters such as coal face spalling and end-face breakage by supplementing non-contact texture information that is difficult to obtain with traditional sensors.

[0042] S5. Construct an intelligent early warning model. Based on historical data, construct a deep learning model that includes an attention mechanism and a long short-term memory network. Use the attention mechanism to dynamically weight and fuse multimodal features, and use the long short-term memory network to capture the temporal evolution pattern and output the top plate to suppress probability and risk level.

[0043] Step S5, the construction and operation of the intelligent early warning model, includes: Feature fusion: The extracted multidimensional feature vectors are input into the attention mechanism module to calculate the importance weight of each feature for predicting roof pressure. And generate the fused feature vector based on the weights. : ; Temporal prediction: The fused feature vector The data is input into a long short-term memory network according to time sequence, and its gating mechanism is used to learn the evolution law of mine pressure.

[0044] Risk Classification: The model outputs the probability value of the roof slab pressing down within a preset time period in the future, and determines the risk level based on the interval to which the probability value belongs.

[0045] The model first utilizes an attention mechanism module to dynamically allocate weights according to different stages of mine pressure evolution, automatically focusing on the most indicative precursory information at the current moment and suppressing non-critical modal interference, significantly improving the accuracy of feature representation. Subsequently, it leverages the gating structure of a long short-term memory network to effectively memorize and filter data over long time spans, solving the problem of capturing long-distance temporal dependencies and achieving deep learning and trend extrapolation of the roof pressure incubation process. Finally, through risk classification, it maps complex predictive probabilities into intuitive early warning levels, providing a clear basis for on-site decision-making and automatically triggering corresponding control logic, realizing rapid transformation from monitoring data to safety management actions.

[0046] The model training process in step S5 includes: A dataset was constructed, and historical monitoring data was divided into non-pressure samples and periodic pressure samples. The periodic pressure samples were further subdivided into weak pressure, medium pressure and strong pressure samples. The model is trained using the pre-defined training set, the loss is calculated using the cross-entropy loss function, and the model parameters are updated using the backpropagation algorithm. Regularization techniques are introduced to randomly discard some network nodes to suppress overfitting, and the model performance is evaluated using a validation set. Training stops when the performance metrics meet the preset requirements.

[0047] This training system first constructs a standardized dataset containing refined intensity levels, providing clear labels for supervised learning. This enables the model to grasp the differentiated characteristics of different levels of mine pressure, significantly improving the precision and targeting of early warnings. In the optimization phase, the cross-entropy loss function is used to quantify distribution differences, and the backpropagation algorithm guides the weights iteratively towards minimizing the loss, achieving rapid convergence of the model's fitting ability. Furthermore, the introduction of regularization techniques effectively reduces the co-adaptation between neurons to enhance robustness. Combined with real-time monitoring of generalization ability using a validation set, the risk of overfitting is avoided, thus establishing the reliability of the final model in applications with unknown samples.

[0048] The specific thresholds for risk classification are set as follows: When the probability of incoming pressure is ≤0.3, it is judged as a low-risk level; When the probability of pressure arrival is less than 0.3 and less than or equal to 0.7, it is classified as a medium-risk level. When the probability of incoming pressure is greater than 0.7, it is classified as a high-risk level. This classification strategy identifies the stability of the surrounding rock by setting a low-risk threshold, effectively shielding false alarms caused by minor disturbances in the rock mass and ensuring continuous and efficient operation of the fully mechanized mining face. At the same time, it uses the medium-risk range to define the transitional stage of mine pressure energy accumulation, providing an early warning buffer period to prompt the strengthening of local monitoring and adjustment of support, preventing untimely response due to sudden incoming pressure. Finally, by locking in the moment of strong incoming pressure with significant disaster risk through the high-risk threshold, it directly triggers the evacuation command or high-resistance support logic, constructing a mandatory safety barrier to prevent large-scale roof collapse.

[0049] S6. Tiered Early Warning and Linked Response: Based on the risk level output by the model, the system automatically triggers corresponding early warning signals and on-site handling measures.

[0050] The graded early warning and linkage response in step S6 includes: When the risk level is determined to be low, the system will display a normal message and will not initiate active alarms. When the risk level is determined to be medium, a voice broadcast and pop-up reminder will be issued, the sensor sampling frequency will be increased, and personnel will be notified to strengthen on-site inspections. When a high-risk level is determined, an audible and visual alarm and a telephone call will be activated. The system will automatically interlock to prevent equipment in the alarm area from performing unnecessary actions, and will forcibly stop operations and evacuate personnel.

[0051] This tiered linkage mechanism employs a normal display mode during low-risk phases, ensuring data transparency while preventing invalid alarms from distracting operators and maintaining normal production order. Once the system enters a medium-risk phase, it alerts personnel through both auditory and visual warnings, increases sensor sampling frequency to prevent the omission of critical precursor information, and effectively eliminates false alarms through manual verification. In high-risk emergencies, high-intensity audible and visual alarms and remote calls are used to forcefully transmit instructions in the noisy underground environment. Simultaneously, automatic interlocking functions are activated to cut off equipment power, blocking the path of mechanical vibration-induced rock instability. Ultimately, through mandatory shutdown and evacuation measures, a robust defense line is constructed, protecting both equipment and personnel.

[0052] S7. Feedback optimization: Compare the actual mining pressure events with the early warning results, and use the comparison results to adjust the parameters of the intelligent early warning model.

[0053] Structure of a multimodal data fusion-based intelligent mine pressure monitoring system: The present invention also provides a multimodal data fusion-based intelligent mine pressure monitoring system, comprising: The data acquisition module is used to drive the distributed fiber optic sensor 1, micro-vibration detector 2, displacement sensor 3, stress sensor 4 and camera 5 to acquire data; The data processing module is configured to perform the data preprocessing and feature extraction steps in step S3 above; The intelligent analysis module has a built-in early warning model based on attention mechanism and long short-term memory network, which is used to perform multimodal feature fusion and pressure probability prediction in step S5 above. The early warning execution module is used to trigger corresponding alarm devices and control commands based on the risk level output by the intelligent analysis module.

[0054] The data acquisition module, serving as the hardware-software interface, achieves synchronous sampling through unified scheduling of heterogeneous sensors, effectively solving the problem of inconsistent time bases for multi-source data and establishing the spatiotemporal correspondence foundation for subsequent analysis. Subsequently, the data processing module acts as a purification and transformation center, removing environmental noise while converting raw data into machine-understandable feature vectors, significantly improving the signal-to-noise ratio and usability of the data. On this basis, the intelligent analysis module, as the core computing unit, delves into the nonlinear correlations between multidimensional features to deduce mine pressure trends, achieving advanced and accurate identification of pressure events and reducing false alarm rates. Finally, the early warning execution module rapidly transforms abstract risk levels into on-site perceptible physical signals and control commands, ensuring the absolute timeliness of emergency response.

[0055] Please see the appendix Figure 1-2 An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor.

[0056] The electronic device uses memory to store system instructions and model parameters, providing stable data read and write support for the operation process and ensuring data integrity after power failure. At the same time, the processor, as the core computing unit, executes complex feature extraction and deep learning inference by calling stored instructions, realizing real-time analysis and rapid response of massive monitoring data. The computer program carries the core algorithm logic, and through deep collaboration between software and hardware, it transforms the multimodal fusion monitoring theory into practically executable automatic control actions, thereby completing the accurate identification of roof pressure risk.

[0057] 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. A multimodal data fusion-based intelligent mining pressure monitoring method, characterized in that, Includes the following steps: S1. Construct a multi-source heterogeneous sensing network. Based on the geological conditions of the mining face and roadway, deploy a variety of monitoring devices including distributed optical fiber sensors (1), microseismic detectors (2), displacement sensors (3), stress sensors (4) and cameras (5). The cameras (5) are installed on hydraulic supports (6) to form a full-space monitoring network covering the mining face and surrounding rock mass. S2. Multimodal data acquisition and transmission: Real-time acquisition of internal stress, vibration signals, displacement data and coal wall fracture images of the rock mass through the sensing network, and transmission of the acquired data to the cloud platform through the edge computing gateway; S3. Data preprocessing: Denoising is performed on data of different modalities, and abnormal or missing data is repaired to obtain standardized time series data. S4. Multidimensional spatiotemporal feature extraction: Extracting temporal features that characterize the precursors of roof pressure from the preprocessed data, including stress statistical features, displacement motion features, microseismic energy features, and crack image features. S5. Construct an intelligent early warning model. Based on historical data, build a deep learning model that includes an attention mechanism and a long short-term memory network. Use the attention mechanism to dynamically weight and fuse multimodal features. Use the long short-term memory network to capture the temporal evolution pattern and output the probability and risk level of the top plate. S6. Tiered early warning and linkage response: Based on the risk level output by the model, the corresponding early warning signal and on-site handling measures are automatically triggered. S7. Feedback optimization: Compare the actual mining pressure events with the early warning results, and use the comparison results to adjust the parameters of the intelligent early warning model.

2. The multimodal data fusion-based intelligent mine pressure monitoring method according to claim 1, characterized in that, The deployment of the sensing network in step S1 specifically includes: Distributed fiber optic sensors (1) are embedded in the roadway roof and coal wall to monitor the stress distribution inside the rock mass. Microseismic detectors (2) are installed on the roof and sidewalls of the two roadways of the working face to collect vibration information of rock mass fracture. Displacement sensors (3) are placed on the roof of the roadway and in both roadways to monitor the displacement of the coal and rock mass and the deformation of the surrounding rock. The stress sensor (4) is installed inside the hydraulic support (6) column and in the tunnel borehole to collect the support pressure and the bearing pressure. The camera (5) is installed below the top beam of the hydraulic support (6), with the lens facing the coal wall, to monitor the development of cracks on the coal wall surface and the contact status between the support and the roof.

3. The multimodal data fusion-based intelligent mine pressure monitoring method according to claim 1, characterized in that, The specific method of data preprocessing in step S3 is as follows: A state-space model is established for stress and displacement data, and Kalman filtering algorithm is used for noise reduction. For microseismic signals, wavelet threshold denoising method is used to remove low-frequency interference components through wavelet decomposition; For the image data, Gaussian filtering is used to remove noise, and histogram equalization is used to adjust the image contrast. The surface areas of the coal wall and roof are extracted through image segmentation. For suspicious or missing data marked in the initial verification, if the deviation is within the preset range, interpolation is used to repair it; if the deviation exceeds the preset range or the data is missing, a time series prediction model is used to supplement the prediction based on historical data trends.

4. The multimodal data fusion-based intelligent mine pressure monitoring method according to claim 1, characterized in that, The feature extraction in step S4 specifically includes: Stress-related characteristics: Calculate the rate of change of stress per unit time, the amplitude of stress fluctuation, and the mean and standard deviation of stress; Displacement-related characteristics: Calculate displacement acceleration, cumulative displacement, and displacement change frequency; Microseismic characteristics: Calculate the number of microseismic events per unit time, total microseismic energy, average energy, and the proportion of high-frequency microseismic events; Image-related features: Cracks on the coal face and roof surface are extracted using image recognition algorithms, and the crack density change rate, average crack length, and average crack width are calculated.

5. The multimodal data fusion-based intelligent mine pressure monitoring method according to claim 1, characterized in that, The construction and operation of the intelligent early warning model in step S5 includes: Feature fusion: The extracted multidimensional feature vector is input into the attention mechanism module, the importance weight of each feature to the prediction of the top plate pressure is calculated, and the fused feature vector is generated according to the weight. Temporal prediction: The fused feature vectors are input into a long short-term memory network according to the time series, and the evolution law of mining pressure is learned by using a gating mechanism; Risk Classification: The model outputs the probability value of the roof slab pressing down within a preset time period in the future, and determines the risk level based on the interval to which the probability value belongs.

6. The multimodal data fusion-based intelligent mine pressure monitoring method according to claim 1, characterized in that, The model training process in step S5 includes: A dataset was constructed, and historical monitoring data was divided into non-pressure samples and periodic pressure samples. The periodic pressure samples were further subdivided into weak pressure, medium pressure and strong pressure samples. The model is trained using the pre-defined training set, the loss is calculated using the cross-entropy loss function, and the model parameters are updated using the backpropagation algorithm. Regularization techniques are introduced to randomly discard some network nodes to suppress overfitting, and the model performance is evaluated using a validation set. Training stops when the performance metrics meet the preset requirements.

7. The multimodal data fusion-based intelligent mine pressure monitoring method according to claim 1, characterized in that, The specific threshold for the risk classification is set as follows: When the probability of pressure is less than or equal to 0.3, it is judged as a low-risk level; When the probability of pressure is greater than 0.3 and less than or equal to 0.7, it is judged as a medium-risk level; When the probability of pressure is greater than 0.7, it is judged as a high-risk level.

8. The multimodal data fusion-based intelligent mine pressure monitoring method according to claim 7, characterized in that, The graded early warning and linkage response in step S6 includes: When the risk level is determined to be low, the system will display a normal message and will not initiate active alarms. When the risk level is determined to be medium, a voice broadcast and pop-up reminder will be issued, the sensor sampling frequency will be increased, and personnel will be notified to strengthen on-site inspections. When a high-risk level is determined, an audible and visual alarm and a telephone call will be activated. The system will automatically interlock to prevent equipment in the alarm area from performing unnecessary actions, and will forcibly stop operations and evacuate personnel.

9. A multimodal data fusion-based intelligent mine pressure monitoring device, comprising the multimodal data fusion-based intelligent mine pressure monitoring method according to any one of claims 1-8, characterized in that, include: The data acquisition module is used to drive the distributed optical fiber sensor (1), micro-vibration detector (2), displacement sensor (3), stress sensor (4) and camera (5) as described in claim 1 to acquire data; The data processing module is configured to perform the data preprocessing and feature extraction steps as described in claim 1 or 3; The intelligent analysis module has a built-in early warning model based on attention mechanism and long short-term memory network, which is used to perform multimodal feature fusion and pressure probability prediction. The early warning execution module is used to trigger corresponding alarm devices and control commands based on the risk level output by the intelligent analysis module.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 8.

Images