Coal mine water disaster identification method and device based on size model fusion

Abstract

The application provides a coal mine water disaster identification method and device based on size model fusion, equipment and storage medium, wherein the method comprises: collecting multi-source water disaster monitoring data and generating a multi-dimensional feature matrix; constructing a vertical large model on the cloud to perform cross-modal feature alignment and global reasoning, and deploying a lightweight small model distilled by knowledge on the edge to perform real-time screening; based on the confidence output by the small model, dynamically triggering cloud-edge collaboration, splicing the global semantic features of the large model with the local detailed features of the small model, and performing dynamic weighted joint reasoning according to the hydrological scene complexity; quantifying the risk level according to the reasoning result and generating an early warning linkage instruction. By constructing a dynamic collaborative architecture of edge small models and cloud large models, the application effectively eliminates the interference of the harsh underground environment, reduces the false positive and false negative rates, meets the demand for limited computing power and real-time response underground, and realizes accurate and early warning of water disasters.

Description

Technical Field

[0001] This application relates to the field of intelligent early warning technology for coal mine water hazards, and in particular to a method, device, equipment and storage medium for identifying coal mine water hazards based on the fusion of large and small models. Background Technology

[0002] Coal mine water hazards are a key hidden danger hindering safe production in coal mines. As coal mining progresses to deeper levels, the coupling effect between the geostress field and the seepage field intensifies, continuously increasing the concealment and suddenness of water hazards. Among related technologies, water hazard prevention and control mainly adopts fixed threshold alarm methods, issuing warnings only after parameters exceed limits, resulting in strong lag and a lack of proactive prevention capabilities. Furthermore, the high dust, high humidity, and strong electromagnetic interference environments underground easily cause sensor data distortion and false alarms, and fixed thresholds cannot be adapted to complex hydrogeological scenarios such as fault fracture zones, significantly reducing the monitoring effectiveness in complex areas. Summary of the Invention

[0003] This application aims to at least partially address one of the technical problems in the related art.

[0004] Firstly, this application proposes a coal mine water hazard identification method based on size model fusion, the method comprising:

[0005] Collect multi-source water hazard monitoring data and perform synchronous alignment preprocessing to generate a multi-dimensional feature matrix dataset; Based on the multi-dimensional feature matrix dataset, a large vertical model for coal mine water hazard identification is constructed, and the large vertical model for coal mine water hazard identification is introduced to train the small water hazard prevention and control model through knowledge distillation. The multi-dimensional feature matrix dataset collected in real time is input into the water hazard prevention and control mini-model deployed at the edge for screening. Based on the target category confidence score output by the water hazard prevention and control mini-model, the coal mine water hazard identification vertical large model in the cloud is triggered to perform collaborative calculation. During the collaborative operation, the global hydrological semantic features output by the vertical large model for coal mine water hazard identification and the local detailed features output by the small model for water hazard prevention and control are extracted, and the features are spliced ​​and fused. The voting weights of the vertical large model for coal mine water hazard identification and the small model for water hazard prevention and control are adjusted in combination with the hydrogeological scene status, and joint inference is performed. The risk level is assessed based on the joint reasoning results, and disaster prevention and control instructions are sent.

[0006] In one implementation, multi-source water hazard monitoring data is collected and synchronously aligned preprocessed to generate a multi-dimensional feature matrix dataset, including: Acquire time-series data including water pressure, water level, water quality, and microseismic data, as well as visual image data and temperature distribution field matrix data; The time series data is processed using a median filtering algorithm combined with a wavelet denoising algorithm, and outlier data points are removed according to the 3σ criterion. The Min-Max normalization method is used to unify the dimensions of the time series data. A global reference timestamp sequence is set. For missing data values ​​at non-reference times, a linear interpolation algorithm is used to calculate the mapping value at the target times, and the time series data of different frequencies are reconstructed into a time series feature matrix with a unified timestamp. The visual image data is processed using a dark channel prior dehazing algorithm and a Retinex image enhancement algorithm. The processed visual image data, the temperature distribution field matrix data, and the temporal feature matrix are then merged to generate the multi-dimensional feature matrix dataset.

[0007] In one implementation, a large-scale vertical model for coal mine water hazard identification is constructed based on the multi-dimensional feature matrix dataset, including: Construct a multimodal feature encoder and generator network for the vertical large-scale model of coal mine water hazard identification; Historical flood damage case records are obtained, and features are extracted from the historical flood damage case records and the multi-dimensional feature matrix dataset by the multimodal feature encoder to generate text feature vectors, image feature vectors and time-series feature vectors. A text-guided dual-branch cross-attention mechanism is adopted, in which the text feature vector is used as the query vector, and dot product calculation and weighted summation are performed with the image feature vector and the temporal feature vector respectively. The output results are then added element-wise to generate the global hydrological semantic features. Using the global hydrological semantic features as query conditions, a search is performed in the preset coal mine hydrological knowledge graph. The generative network, combined with the retrieved associated graph knowledge, performs autoregressive decoding to generate inference results, thus completing the construction of the vertical large model for coal mine water hazard identification.

[0008] In one implementation, the large-scale vertical model for coal mine water hazard identification is introduced to train the small-scale water hazard prevention and control model using knowledge distillation, including: The large vertical model for coal mine water hazard identification is used as the teacher model, and the small water hazard prevention and control model is used as the student model. The category probability distribution output by the teacher model is extracted as a soft label, and a feature adaptation layer is cascaded between the corresponding intermediate feature extraction layers of the teacher model and the student model. The weight parameters of the water hazard prevention and control mini-model are updated iteratively by combining the soft label distribution divergence loss, the L2 loss between the intermediate feature extraction layers, and the cross-entropy loss of the hard labels of the real samples; The water hazard prevention and control mini-model that has completed knowledge distillation training is subjected to network quantization and structured pruning operations.

[0009] In one implementation, based on the target category confidence score output by the small-scale water hazard prevention and control model, the cloud-based vertical large-scale coal mine water hazard identification model is triggered to perform collaborative computation, including: Set a first preset threshold and a second preset threshold, wherein the first preset threshold is greater than the second preset threshold; When the confidence level of the target category is greater than or equal to the first preset threshold, the judgment result of the water hazard prevention and control mini-model is used as the final identification result. When the confidence level of the target category is within the range of the second preset threshold to the first preset threshold, the multi-dimensional feature matrix dataset collected in real time is uploaded to the cloud server, triggering the collaborative operation of the coal mine water hazard identification vertical large model.

[0010] In one implementation, the global hydrological semantic features output by the large-scale vertical model for coal mine water hazard identification are extracted and the local detailed features output by the small-scale water hazard prevention model are concatenated and fused, including: The global hydrological semantic features and the local detail features are concatenated and cascaded along the channel dimension to form an initial joint feature; The initial joint features are input into the channel attention network module, and the gating weight coefficients corresponding to each feature channel are generated through a trainable weight matrix and a nonlinear activation function. The initial joint features are element-wise weighted and mapped using the gating weight coefficients to output a fused feature vector.

[0011] In one implementation, the voting weights of the large-scale vertical model for coal mine water hazard identification and the small-scale water hazard prevention model are adjusted based on the hydrogeological scene conditions, including: Analyze the current hydrogeological scene status of the working face; When the hydrogeological scene is determined to be in a fault intersection zone, a water-rich aquifer zone, or a fault fracture zone, the voting weight of the large vertical model for coal mine water hazard identification is increased and the voting weight of the small water hazard prevention and control model is decreased. When the hydrogeological scene is determined to be a normal single stratum or other conventional hydrogeological scene, the voting weight of the coal mine water hazard identification vertical large model is reduced and the voting weight of the water hazard prevention and control small model is increased.

[0012] In one implementation, after performing the joint inference, the method further includes: The multi-class probability set output by the vertical large model for coal mine water hazard identification and the multi-class probability set output by the small model for water hazard prevention and control are horizontally concatenated by feature dimension to construct a secondary meta-feature matrix. Using the secondary meta-feature matrix as input data and the manually verified real water hazard category results as supervision labels, a meta-classifier based on logistic regression algorithm is trained to establish the mapping relationship between the secondary meta-feature matrix and the posterior probability of the final water hazard category, thus fixing the decision boundary in the dual-end model fusion state.

[0013] In one implementation, the risk level is assessed based on the joint inference result, and a disaster prevention linkage control command is sent, including: Based on the joint reasoning results, the highest probability value of the corresponding water hazard category is obtained, a three-dimensional risk assessment index system covering probability, severity and exposure degree is established, the relative weights of probability, severity and exposure degree are calculated using the analytic hierarchy process, the scores of each standardized index are multiplied by the corresponding relative weights and summed to calculate the comprehensive risk index, and the graded risk level is evaluated based on the set threshold range in which the comprehensive risk index is located. Based on the joint reasoning results and the risk level, a structured early warning message is generated, which includes the three-dimensional spatial location of the hidden danger, the type of hidden danger, the risk level and the handling suggestions, and the structured early warning message is pushed to the terminal. When the risk level in the structured early warning message reaches the set high-risk warning level, a control protocol frame is sent to the downhole programmable logic controller through the communication interface to activate the drainage equipment and lock the mining equipment associated with the three-dimensional spatial location of the hidden danger. The system receives feedback on the actual physical condition of potential hazards after manual review and updates the dynamic sample library. When the set update cycle arrives, it extracts newly labeled data from the dynamic sample library to perform online iterative fine-tuning training on the large vertical model for coal mine water hazard identification and the small model for water hazard prevention and control.

[0014] Secondly, this application proposes a coal mine water hazard identification device based on size model fusion, the device comprising: The first processing module is used to collect multi-source water hazard monitoring data and perform synchronous alignment preprocessing to generate a multi-dimensional feature matrix dataset. The second processing module is used to construct a large vertical model for coal mine water hazard identification based on the multi-dimensional feature matrix dataset, and to use the large vertical model for coal mine water hazard identification to train the small water hazard prevention and control model through knowledge distillation. The third processing module is used to input the real-time collected multi-dimensional feature matrix dataset into the water hazard prevention and control small model deployed at the edge for screening, and trigger the coal mine water hazard identification vertical large model in the cloud to perform collaborative calculation based on the target category confidence output by the water hazard prevention and control small model. The fourth processing module is used to extract the global hydrological semantic features output by the vertical large model for coal mine water hazard identification and the local detailed features output by the small model for water hazard prevention and control during the collaborative operation, perform feature splicing and fusion, and adjust the voting weights of the vertical large model for coal mine water hazard identification and the small model for water hazard prevention and control in combination with the hydrogeological scene status, and perform joint inference. The fifth processing module is used to assess the risk level based on the joint reasoning results and send disaster prevention and control instructions.

[0015] In one implementation, the first processing module can be used to: Acquire time-series data including water pressure, water level, water quality, and microseismic data, as well as visual image data and temperature distribution field matrix data; The time series data is processed using a median filtering algorithm combined with a wavelet denoising algorithm, and outlier data points are removed according to the 3σ criterion. The Min-Max normalization method is used to unify the dimensions of the time series data. A global reference timestamp sequence is set. For missing data values ​​at non-reference times, a linear interpolation algorithm is used to calculate the mapping value at the target times, and the time series data of different frequencies are reconstructed into a time series feature matrix with a unified timestamp. The visual image data is processed using a dark channel prior dehazing algorithm and a Retinex image enhancement algorithm. The processed visual image data, the temperature distribution field matrix data, and the temporal feature matrix are then merged to generate the multi-dimensional feature matrix dataset.

[0016] In one implementation, the second processing module can be used for: Construct a multimodal feature encoder and generator network for the vertical large-scale model of coal mine water hazard identification; Historical flood damage case records are obtained, and features are extracted from the historical flood damage case records and the multi-dimensional feature matrix dataset by the multimodal feature encoder to generate text feature vectors, image feature vectors and time-series feature vectors. A text-guided dual-branch cross-attention mechanism is adopted, in which the text feature vector is used as the query vector, and dot product calculation and weighted summation are performed with the image feature vector and the temporal feature vector respectively. The output results are then added element-wise to generate the global hydrological semantic features. Using the global hydrological semantic features as query conditions, a search is performed in the preset coal mine hydrological knowledge graph. The generative network, combined with the retrieved associated graph knowledge, performs autoregressive decoding to generate inference results, thus completing the construction of the vertical large model for coal mine water hazard identification.

[0017] In one implementation, the second processing module can be used for: The large vertical model for coal mine water hazard identification is used as the teacher model, and the small water hazard prevention and control model is used as the student model. The category probability distribution output by the teacher model is extracted as a soft label, and a feature adaptation layer is cascaded between the corresponding intermediate feature extraction layers of the teacher model and the student model. The weight parameters of the water hazard prevention and control mini-model are updated iteratively by combining the soft label distribution divergence loss, the L2 loss between the intermediate feature extraction layers, and the cross-entropy loss of the hard labels of the real samples; The water hazard prevention and control mini-model that has completed knowledge distillation training is subjected to network quantization and structured pruning operations.

[0018] In one implementation, the third processing module can be used for: Set a first preset threshold and a second preset threshold, wherein the first preset threshold is greater than the second preset threshold; When the confidence level of the target category is greater than or equal to the first preset threshold, the judgment result of the water hazard prevention and control mini-model is used as the final identification result. When the confidence level of the target category is within the range of the second preset threshold to the first preset threshold, the multi-dimensional feature matrix dataset collected in real time is uploaded to the cloud server, triggering the collaborative operation of the coal mine water hazard identification vertical large model.

[0019] In one implementation, the fourth processing module can be used for: The global hydrological semantic features and the local detail features are concatenated and cascaded along the channel dimension to form an initial joint feature; The initial joint features are input into the channel attention network module, and the gating weight coefficients corresponding to each feature channel are generated through a trainable weight matrix and a nonlinear activation function. The initial joint features are element-wise weighted and mapped using the gating weight coefficients to output a fused feature vector.

[0020] In one implementation, the fourth processing module can be used for: Analyze the current hydrogeological scene status of the working face; When the hydrogeological scene is determined to be in a fault intersection zone, a water-rich aquifer zone, or a fault fracture zone, the voting weight of the large vertical model for coal mine water hazard identification is increased and the voting weight of the small water hazard prevention and control model is decreased. When the hydrogeological scene is determined to be a normal single stratum or other conventional hydrogeological scene, the voting weight of the coal mine water hazard identification vertical large model is reduced and the voting weight of the water hazard prevention and control small model is increased.

[0021] In one implementation, after performing the joint inference, the method further includes: The multi-class probability set output by the vertical large model for coal mine water hazard identification and the multi-class probability set output by the small model for water hazard prevention and control are horizontally concatenated by feature dimension to construct a secondary meta-feature matrix. Using the secondary meta-feature matrix as input data and the manually verified real water hazard category results as supervision labels, a meta-classifier based on logistic regression algorithm is trained to establish the mapping relationship between the secondary meta-feature matrix and the posterior probability of the final water hazard category, thus fixing the decision boundary in the dual-end model fusion state.

[0022] In one implementation, the fifth processing module can be used for: Based on the joint reasoning results, the highest probability value of the corresponding water hazard category is obtained, a three-dimensional risk assessment index system covering probability, severity and exposure degree is established, the relative weights of probability, severity and exposure degree are calculated using the analytic hierarchy process, the scores of each standardized index are multiplied by the corresponding relative weights and summed to calculate the comprehensive risk index, and the graded risk level is evaluated based on the set threshold range in which the comprehensive risk index is located. Based on the joint reasoning results and the risk level, a structured early warning message is generated, which includes the three-dimensional spatial location of the hidden danger, the type of hidden danger, the risk level and the handling suggestions, and the structured early warning message is pushed to the terminal. When the risk level in the structured early warning message reaches the set high-risk warning level, a control protocol frame is sent to the downhole programmable logic controller through the communication interface to activate the drainage equipment and lock the mining equipment associated with the three-dimensional spatial location of the hidden danger. The system receives feedback on the actual physical condition of potential hazards after manual review and updates the dynamic sample library. When the set update cycle arrives, it extracts newly labeled data from the dynamic sample library to perform online iterative fine-tuning training on the large vertical model for coal mine water hazard identification and the small model for water hazard prevention and control.

[0023] Thirdly, this application provides an electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method as described in the first aspect.

[0024] Fourthly, this application proposes a storage medium storing instructions that, when executed on an electronic device, cause the electronic device to perform the method described in the first aspect.

[0025] Fifthly, this application proposes a program product comprising at least one of a program and instructions, wherein when the program and instructions are executed by an electronic device, they implement the steps of the method described in the first aspect.

[0026] The coal mine water hazard identification method, device, equipment, and storage medium provided in this application, based on the fusion of large and small models, can construct a multi-dimensional feature matrix dataset by synchronously aligning and preprocessing multi-source water hazard monitoring data, build a vertical large model for coal mine water hazard identification, and optimize and train a lightweight water hazard prevention and control small model at the underground edge through knowledge distillation. Relying on a cloud-edge collaborative architecture, the edge small model performs preliminary screening of real-time monitoring data, and intelligently triggers collaborative computation of the cloud large model based on the output confidence level. It integrates the global hydrological semantic features of the large model with the local detailed features of the small model, and dynamically adjusts the voting weights of the two models in combination with the on-site hydrogeological scene to complete joint inference, obtain the water hazard risk level, and issue disaster prevention and control instructions. This solution can effectively eliminate the monitoring impact caused by the harsh underground environment, while adapting to the limited computing power of underground edge equipment. It solves the defects of weak feature extraction capability of single model, poor adaptability to complex scenarios, and low sensitivity of early anomaly identification, and significantly improves the intelligent prevention and control level of water hazards in deep and complex mines.

[0027] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0028] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a coal mine water hazard identification method based on size model fusion provided in this application; Figure 2 This is a schematic diagram of the structure of a coal mine water hazard identification device based on size model fusion provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0029] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0030] The coal mine water hazard identification method and apparatus based on size model fusion according to embodiments of this application are described below with reference to the accompanying drawings.

[0031] See attached document Figure 1 , Figure 1 This is a flowchart illustrating a coal mine water hazard identification method based on size model fusion provided in this application, as shown below. Figure 1 As shown, the method may include the following steps: S100 collects downhole multi-source water hazard monitoring data and performs filtering, noise reduction, and unified alignment preprocessing to generate a multi-dimensional dataset that conforms to the input specifications. S200 combines a coal mine hydrological knowledge graph to construct a large vertical model for coal mine water hazard identification. Through a multimodal feature encoder, different types of data are mapped to a unified semantic space, and a cross-modal joint reasoning mechanism is established. S300 is based on a multi-task learning framework to build a small model for water hazard prevention and control. It introduces a vertical large model to perform knowledge distillation to transfer soft labels through a feature adaptation layer, and performs quantization and pruning operations on the output parameters to generate a lightweight network. S400 dynamically triggers vertical large model collaborative operation based on confidence threshold when the small model of water hazard prevention and control performs edge real-time screening. It extracts global hydrological semantic features of the large model and local detailed features of the small model and performs splicing. It adjusts the model judgment weights and performs joint inference based on the complexity of the hydrological scene. The S500 classifies risk levels and generates structured early warning information based on the joint reasoning results, sends disaster prevention linkage control commands to the downhole programmable logic controller, and performs model training iteration based on on-site verification data.

[0032] In this embodiment, step S10 may include the following steps: S101, acquire physical environmental parameters and hydrological characteristic signals of the mining area.

[0033] For example, intelligent water pressure and level sensors are deployed in key aquifer areas where Ordovician limestone water and old workings water are located. As a preferred approach, the sensor's acquisition frequency is configured with a dual-mode system of timed acquisition and event-triggered acquisition. The timed acquisition frequency is set to execute once every five minutes. When the water pressure fluctuation exceeds a preset water pressure change threshold (e.g., 0.1 MPa), it automatically switches to event-triggered mode for high-frequency recording to capture drastic hydrodynamic changes before a water hazard occurs, maintaining a monitoring accuracy within 0.01 MPa. The water inflow metering device in the sensor data acquisition module is installed at drainage nodes, collecting time-series data on the total mine inflow and segmented inflows at a sampling frequency of once per minute. Water quality sensors collect data on calcium ion concentration, sulfate ion concentration, and pH value in the water. The normal water quality acquisition frequency is once per hour, and the sampling frequency is automatically increased when the ion concentration change rate exceeds a preset calibration time. Microseismic monitoring equipment is distributed within the rock mass, collecting weak vibration signals generated by rock fractures to identify the development status of water-conducting fractures, with a spatial positioning accuracy of less than or equal to 10 meters. The camera video acquisition module includes a high-definition camera and an infrared thermal imaging device. This module is deployed in low-lying areas of underground tunnels, aquifer outcrops, and fault fracture zones. The high-definition camera acquires a visual image stream with a resolution of at least 1920×1080 at a frame rate of 25 frames per second, and enables low-light enhancement to record images of water accumulation, rock wall seepage, and piping in real time. The infrared thermal imaging device simultaneously acquires temperature distribution field matrix data of the rock surface to help identify the location of abnormal water-rich structures.

[0034] S102, synchronously extract geological features and production space correlation data.

[0035] Specifically, considering that the occurrence of water hazards is not only related to transient hydrological changes but also constrained by underlying geological structures, three-dimensional hydrogeological model data of the mining area can be used to extract fault location coordinates, collapse column distribution boundaries, aquifer thickness values, and permeability physical parameters in three-dimensional space. This data can then be integrated with coal mine production execution to extract corresponding mining face advance speed, mining height data, and support parameters, constructing a database linking geology, production, and water hazards. The hydrological data interface synchronously aggregates historical water hazard case records, establishing a feature set encompassing geological information, production parameters, and water hazard characteristics.

[0036] S103 performs noise reduction, anomaly removal, and timestamp interpolation alignment calculations on the acquired multi-source heterogeneous sensor signals to generate a standardized time series matrix.

[0037] It should be noted that in the complex and highly electromagnetically interfered environment of underground coal mines, sensors are prone to generating spike noise signals. A median filtering algorithm combined with wavelet denoising is used to process the sensor's time-series data, removing high-frequency physical noise generated by the operation of the equipment. After basic denoising, outlier data points caused by sudden environmental changes are identified and removed based on the 3σ criterion. Specifically, the expected value and standard deviation of the sensor measurements are calculated in real time based on a sliding time window (e.g., multiple consecutive historical sampling periods for a water pressure sensor). When the absolute difference between the current sampling point's measurement and the expected value exceeds three times the standard deviation, the measurement is marked as an outlier and removed, thus avoiding interference from extreme erroneous values ​​in subsequent model training.

[0038] To address the issue of inconsistent data dimensions among different types of sensors, the Min-Max normalization method can be used to map one-dimensional time-series data to the [0,1] interval. In practice, the maximum and minimum values ​​are extracted from the current feature sample set or a preset historical calibration dataset. The difference between the current measurement and the minimum value is calculated and divided by the range between the maximum and minimum values. This scaling mapping eliminates the interference of different physical dimensions such as water pressure and water quality on subsequent neural network weight updates.

[0039] Since the hardware sampling clock and reporting period of multi-source sensor devices differ, a linear interpolation algorithm can be used to achieve unified timestamp alignment of multi-dimensional time-series data. A global reference timestamp sequence is established. For missing data values ​​at non-reference time points, the two adjacent original valid sampling timestamps and their corresponding acquired values ​​before and after the target time point are found. Based on the slope of change between these two adjacent points, the mapping value of the target time point is calculated, and boundary condition checks are performed to strictly limit the target time point to be within the closed interval formed by the two original valid timestamps, thus avoiding non-convergent extrapolation calculations. Through this process, discrete sensor data of different frequencies are reconstructed into a multi-dimensional time-series feature matrix with a unified sampling period.

[0040] S104, perform environment-adaptive enhancement processing on the acquired visual image.

[0041] To address the optical feature attenuation caused by the high dust and low illumination environment in underground mines and reduce interference with subsequent image encoders, a dark channel prior dehazing algorithm is used to process the dust fogging effect in the image, and the Retinex image enhancement algorithm is applied to adjust the local contrast and illumination reflection distribution of the image. For the specific derivation and calculation process of the dark channel prior dehazing algorithm and the Retinex algorithm, those skilled in the art can refer to existing digital image processing literature for conventional implementation; their specific implementation processes are well-known in the field and will not be elaborated here. After the aforementioned sensor signal interpolation and image enhancement processing, a multi-dimensional feature matrix dataset suitable for the input specifications of subsequent deep learning network models is generated.

[0042] In this embodiment, performing feature unification and logical reasoning based on a multi-dimensional feature matrix dataset may include the following steps: S201 deploys the infrastructure for large industry-specific models on cloud servers and performs targeted fine-tuning training.

[0043] To ensure the extraction capability of deep semantic features, a pre-trained language model based on the Transformer decoder architecture was selected as the base network. As a preferred implementation, the Sunstone Mine large-scale model with hundreds of billions of parameters was chosen as the base network to retain its powerful basic semantic understanding capabilities. To adapt the general base model to the field of coal mine hydrogeology, a specialized fine-tuning dataset was constructed. The samples in this dataset are derived from historical water inrush accident investigation reports, geological exploration borehole data, and daily inspection records. The dataset labels are set as expert-annotated water hazard types and their corresponding four-level risk levels. During the model fine-tuning training phase, a low-rank adaptive fine-tuning technique was applied. This technique, while keeping the original deep network weight parameters of the base model frozen, injects trainable dimensionality-reduced and dimensionality-increasing linear matrices into the side path of the self-attention layer of the Transformer architecture. The cross-entropy loss function is used to calculate the error between the model's predicted output and the true label, and the initial learning rate is configured within the conventional range of 0.01% to 0.05%. The weight parameters of the side path matrix are updated separately through the backpropagation algorithm. Through this training, the model can establish a deep mapping relationship between professional terminology for coal mine seepage prevention and water control and reasoning logic with relatively low computational resource overhead.

[0044] S202, the vertical large model maps different types of data to a unified high-dimensional semantic space through a multimodal feature encoder, and establishes a cross-modal joint reasoning mechanism.

[0045] The input of the vertical large-scale model is configured with three parallel feature extraction networks: a text encoder, an image encoder, and a temporal encoder. The text encoder uses a BERT-based network structure, receiving geological report text and historical case records, extracting syntactic and context-dependent features, and projecting the feature vectors into a pre-defined 768-dimensional text feature vector through fully connected layers. The image encoder uses a ViT-based network structure, receiving pre-processed downhole visual images and infrared temperature field matrices. The image encoder segments the input 2D image into a fixed-size image patch sequence, extracts local texture and global contour features through multi-layer self-attention modules, and outputs a 768-dimensional image feature vector through a linear mapping layer. The temporal encoder uses a Bi-LSTM (Bidirectional Long Short-Term Memory) network structure, receiving multi-dimensional temporal feature matrices such as water pressure, water quality, and microseismic data. The temporal encoder captures forward and backward temporal dependencies to identify hydrodynamic abrupt changes, and its final hidden state is also mapped to a 768-dimensional temporal feature vector.

[0046] After obtaining the feature vectors of the three modalities mentioned above, the vertical large model introduces a cross-modal attention fusion module to perform deep semantic alignment. To avoid dimensionality conflicts between different modalities, this module adopts a text-guided dual-branch cross-attention mechanism.

[0047] Specifically, the 768-dimensional text feature vector can be used as a shared query vector. In the first branch, the image feature vector is used as both the first key vector and the first value vector to perform a dot product calculation to obtain the first attention weight distribution. This weight distribution is then applied to the first value vector for weighted summation, thereby outputting text-guided image semantic features. Similarly, in the second branch, the temporal feature vector is used as both the second key vector and the second value vector to perform the same dot product and weighted summation calculation, outputting text-guided temporal semantic features. Subsequently, the output image semantic features and temporal semantic features are fused element-wise. Through the above calculations, unstructured image visual features and structured sensor numerical features are forcibly aligned to a unified space guided by text semantics, outputting a fused 768-dimensional global hydrological semantic feature vector. This vector comprehensively represents the multi-dimensional water hazard development status of the current working face.

[0048] S203, a vertical large model combined with a coal mine hydrological knowledge graph to initiate retrieval and enhance generative reasoning, outputting global analysis results with interpretability.

[0049] For example, a professional knowledge graph of coal mine water hazards is pre-constructed. Entity nodes in the graph are defined as disaster-causing factors, hydrological anomalies, and treatment measures, while edges between entities are defined as causal relationships and spatial topological relationships. The graph data is converted into vectors and stored in a high-dimensional vector database in the cloud. During online inference, the vertical large model uses the 768-dimensional global hydrological semantic feature vector generated by S202 as query conditions and calculates cosine similarity in the vector database. Historical similar water inrush case nodes with recall similarity scores higher than a preset threshold (usually set between 0.80 and 0.95 to ensure high relevance of recalled cases) and their corresponding standard treatment plan subgraphs are retrieved. The vertical large model's generative network concatenates the recalled structured graph knowledge with the current multimodal data input and performs autoregressive decoding generation. Finally, the vertical large model outputs risk level judgment data representing the probability of water hazard classification to the fusion layer, and simultaneously outputs structured inference process description text and specific grouting and water plugging or evacuation and shutdown suggestions. This mechanism can effectively improve the black-box unexplainable problem of deep learning networks and provide on-site managers with technically traceable disaster prevention decision-making basis.

[0050] In this embodiment, the model layer focuses on building a high-performance small model adapted for coal mine water hazard identification and assessment using edge computing devices. This model combines a multi-task learning framework with knowledge distillation techniques to address the insufficient feature extraction capabilities of lightweight networks, specifically including the following steps: S301 is a multi-task feature network built on a lightweight convolutional neural network architecture for edge computing.

[0051] To adapt to the computational resource constraints of the underground explosion-proof industrial control computer, a lightweight network based on the MobileNetV3 architecture was selected as the backbone structure of the small-scale water hazard prevention and control model. As a preferred approach, this backbone structure can also be replaced with ShuffleNetV2 or EfficientNet-Lite architectures depending on the instruction set differences of the edge chips. This network architecture uses depthwise separable convolutional layers to replace standard spatial convolution operations, decoupling conventional spatial dimension convolutions from channel dimension convolutions, thereby effectively reducing the number of parameters in the computation process. To enhance the ability to capture key features, the SE-Net channel attention mechanism is introduced into the backbone network, assigning adaptive weights to feature maps of different channels to enhance the model's attention to key features such as sudden changes in underground water pressure and seepage textures on the rock wall. Simultaneously, a multi-scale feature fusion structure is constructed within the network, splicing and fusing extracted shallow edge detail features with deep global semantic features to improve the ability to identify water hazard hazards at different spatial scales.

[0052] Based on the aforementioned multi-task learning framework, the water hazard prevention and control mini-model configures specialized feature learning branches and corresponding input data feature sets for different water hazard types. Specifically, for the old workings water inrush identification branch, the defined input data includes water pressure surge values, abnormal water quality ion concentration indicators, microseismic signal frequency characteristics, and corresponding mining location data; for the Ordovician limestone water inrush identification branch, its input data includes the long-term trend sequence of aquifer water pressure changes, microseismic features characterizing the development of water-conducting fractures, and geological structure data; for the surface water intrusion identification branch, the input data includes the time series of wellhead water level changes, visual characteristics of the spread speed of water accumulation in tunnels, and meteorological and rainfall data. After feature forward propagation, the output layer of the water hazard prevention and control mini-model simultaneously outputs results with three specific physical meanings: the type of water hazard hazard warning, the corresponding risk level, and the location coordinates of the water hazard occurrence, to meet the needs of precise on-site prevention and control.

[0053] S302 constructs a cross-network layer feature adaptation and knowledge distillation learning mechanism, using a vertical large model to guide the parameter optimization of a small model for water hazard prevention and control.

[0054] During the training phase, a fully constructed industry-specific large model is used as the teacher model, and a lightweight small model is used as the student model. Model training employs a soft-label learning strategy, extracting the class probability distribution output by the teacher model under the same input data as soft labels, guiding the student model to fit this probability distribution rather than simply fitting the one-hot encoded hard labels. In addition to probability alignment at the output layer, a cascaded feature adaptation layer is placed between the corresponding intermediate feature extraction layers of the teacher and student models. By calculating the L2 loss function between the feature vectors of these intermediate layers, structured transfer of deep implicit knowledge is achieved.

[0055] During the overall model update process, the weights of the small model are updated iteratively by combining the soft label distribution divergence loss, the intermediate layer L2 loss, and the cross-entropy loss of the hard labels of the real samples. Through the above knowledge transfer process, the distilled water hazard prevention and control small model can achieve more than 90% of the basic performance of the vertical large model in core indicators such as recognition accuracy.

[0056] S303 performs edge-side network quantization, pruning optimization, and inference pipeline reconstruction on the water hazard control mini-model that has completed distillation training.

[0057] To meet the extremely low latency requirements of edge devices, structured pruning techniques are applied to evaluate network channel weights. Under the boundary condition of less than 1% loss in model recognition accuracy, redundant network branches with low contribution to the output are pruned, reducing the overall model parameter count by 30% to 50%. After structured pruning, low-bit quantization of the network weights is performed, mapping the model's internal weight parameters and activation values ​​from 32-bit floating-point (FP32) data format to 8-bit integer (INT8) data format. This conversion reduces the model's physical storage space footprint by approximately 75% and directly adapts to the underlying hardware's integer acceleration unit.

[0058] During the deployment phase, the reconstructed water hazard prevention and control mini-model is encapsulated and invoked using the TensorRT inference engine at the edge device. Batch processing and asynchronous inference techniques are employed to optimize the data input / output pipeline. For the memory allocation and parallel computing scheduling principles of the TensorRT inference engine, those skilled in the art can refer to existing literature on deep learning acceleration deployment technologies for conventional configuration; the specific computational logic is well-known in the field and will not be elaborated upon here. Through the aforementioned pruning, quantization, and engine optimization, the inference latency for a single frame image or a single temporal feature matrix at the edge is ensured to remain within 100 milliseconds, meeting the stringent timeliness requirements for real-time monitoring of water hazards in coal mines. In this embodiment, a scene-aware, weight-adaptive, and dual-end collaborative inference system is established. The model fusion strategy is dynamically adjusted according to the complexity of the hydrological environment to achieve reasonable allocation of computing resources for recognition accuracy and real-time response. This may include the following steps: S401 deploys an adaptive collaborative reasoning process and an edge-cloud routing and distribution mechanism.

[0059] For example, a quantified and pruned small-scale water hazard prevention model is deployed on edge computing device nodes such as downhole cameras and sensor gateways. This edge network continuously performs forward propagation calculations on the preprocessed real-time multi-dimensional feature matrix, outputting an initial confidence level indicating whether the current environmental state is in the normal or suspected hazard category. To improve resource utilization and avoid cloud computing overload, a hierarchical data routing rule based on confidence level is established: When the confidence level of the target category output by the small model is greater than or equal to 0.9, it is determined that the current input feature is completely within the representation capability coverage range of the small model, and it is directly output as the final recognition result. The edge device directly triggers an early warning command, and the vertical large model in the cloud does not need to call computing power for analysis. When the confidence level of the target category output by the small model is distributed between 0.5 and 0.9, it is determined that the current feature has boundary ambiguity, and the cross-end collaboration mechanism is activated to upload the multi-source raw data flow of the corresponding time period to the cloud server, triggering the vertical large model to perform deep analysis. When the confidence level of the highest category output by the small model is lower than 0.5, it is determined to be normal environmental background noise or a state without potential risks, and only local log storage is performed without triggering any alarm or data reporting actions. For the data flow control allocation and message queue management logic at the network communication layer, those skilled in the art can use the existing edge IoT transmission protocol stack for conventional configuration. Its specific implementation mechanism is a well-known technology in this field and will not be described in detail here.

[0060] S402 constructs a cross-scale feature-level stitching and attention-weighted fusion computing flow in the cloud to achieve physical alignment of global hydrological semantic features and local detailed features.

[0061] For example, when cloud collaboration is triggered, the large vertical model extracts global hydrological semantic features reflecting the overall hydrogeological conditions, fault zone distribution, and mining impact range of the mining area. These semantic features are then fixed as a 1024-dimensional feature vector through an internal linear projection layer. The hidden layer output of the small water hazard prevention model, simultaneously uploaded to the cloud, carries transient change information such as sudden changes in water pressure and rock wall seepage textures, and is extracted as a 512-dimensional local detail feature vector. Since the macroscopic features extracted by the large model and the microscopic features extracted by the small model differ in their representational scale, directly adding the values ​​can easily obscure key signals. Therefore, as a preferred approach, feature concatenation and an attention mechanism are used for fusion. Specifically, the 1024-dimensional feature vector output by the large model and the 512-dimensional feature vector output by the small model are concatenated along the channel dimension to form an initial 1536-dimensional joint feature. This initial joint feature is then input into the channel attention network module, where a trainable weight matrix and a nonlinear activation function are used to generate gating weight coefficients for each feature channel. The initial 1536-dimensional joint features are element-wise weighted by the weight coefficient, thereby adaptively amplifying the key representation dimensions under the current water disaster scenario and outputting the final fused feature vector.

[0062] S403 implements a decision-level weighted voting fusion and adaptively schedules the weight allocation scheme based on the underlying hydrogeological scene status.

[0063] For example, at the decision output end, the large model branch outputs discrete risk level judgment values ​​and disaster type results based on graph semantic reasoning; the small model branch outputs tensor distributions containing the classification probabilities of seepage, inrush, sudden water flow, and collapse, as well as the three-dimensional coordinates of the location. To balance expert knowledge and field observation, a dynamic weighting rule is used to synthesize the probability matrices of both ends. The geological structure identifiers in the input matrix are analyzed. When the current working face is perceived to be in a complex scenario such as a fault intersection zone, a water-rich aquifer zone, or a fault fracture zone, the decision intervention degree of global semantic features is increased, setting the voting weight of the large model to 0.7 and the voting weight of the small model to 0.3. When the working face is identified as being in a normal single stratum or other conventional hydrogeological scenario, the system increases the sensitivity priority of local abrupt change signals, reducing the voting weight of the large model to 0.3 and increasing the voting weight of the small model to 0.7. The system allows the execution flow when the logical judgments of the two models are consistent. If the hazard types identified by the large model and the small model are inconsistent, they will be marked as abnormal conflict items and pushed to the pipeline for manual review. To ensure production safety, during the manual review, preventive prompts will be sent to the site based on the transient early warning results of the small model to ensure the accuracy and timeliness of the decision-making terminal.

[0064] S404 trains a meta-classifier based on a Stacking ensemble learning architecture to fix the decision boundary in the fusion state of the two-end models.

[0065] For example, during the optimization training cycle of the fusion strategy, the multi-class probability set output by the large vertical model and the multi-class probability set output by the small water hazard control model are horizontally concatenated by feature dimension to construct a secondary meta-feature matrix. Assuming that both the large and small models output predicted probabilities for four types of water hazards, this secondary meta-feature matrix is ​​an 8-dimensional feature vector. Using this secondary meta-feature matrix as input data and manually verified real water hazard category results as supervision labels, a meta-classifier based on the logistic regression algorithm is trained in the cloud. In the specific training implementation, the meta-classifier establishes a mapping relationship between the secondary meta-feature matrix and the final posterior probability of the water hazard category by iteratively updating the internal weight coefficient matrix and bias term, and uses the cross-entropy loss function to calculate the difference between the predicted probability and the real label. In the validation set stage, the convergence condition of the objective function is to maximize the F1-score (i.e., the harmonic mean of precision and recall). A grid search algorithm is used to traverse the hyperparameter space to lock the weight parameter combination that makes the F1-score reach its peak, thereby improving the generalization performance of the fusion decision model in complex scenarios.

[0066] In this embodiment, based on the fusion features and decision boundaries output by the aforementioned steps, risk quantification assessment of potential water hazards, multi-dimensional control command distribution, and online iterative upgrade of the model status can be performed to form a complete technical chain of identification, early warning, handling, and optimization, which may include the following steps.

[0067] S501 uses a classification network to output the physical classification results of water hazard risks and maps them to the initial severity level.

[0068] For example, after receiving the joint decision vector output by the aforementioned decision fusion module (i.e., the dynamically weighted or stacking meta-classifier), it is input into the final classification module. Specifically, this module is configured with a fully connected layer, which projects the fused joint decision vector onto the specific feature space of the corresponding classification category through a linear transformation of the weight matrix. Subsequently, the network accesses a Softmax multi-classifier for exponential normalization processing, outputting specific probability values ​​for four categories of results: the current monitoring area is in a normal state, there is a risk of old-age water intrusion, there is a risk of Ordos limestone water intrusion, and there is a risk of surface water intrusion. The category with the highest probability value is selected as the current hazard type identification result. Furthermore, based on the numerical range of the highest probability value, the hazard development stage is mapped into three severity levels: initial signs, abnormality, and critical water inrush, providing a basic prerequisite for subsequent comprehensive risk quantification.

[0069] S502 combines multi-dimensional features to construct a risk assessment indicator system, calculates a comprehensive risk index, and triggers corresponding early warning thresholds.

[0070] For example, considering that the destructiveness of underground water hazards in coal mines is influenced by a combination of physical factors, a three-dimensional risk assessment index system encompassing probability, severity, and exposure level can be established. The Analytic Hierarchy Process (AHP) is then used to calculate the relative weights of these three assessment dimensions. In the specific quantitative assessment execution logic, the standardized scores of these three indicators are obtained, multiplied by their corresponding relative weights, and summed to output a comprehensive risk index normalized to the range of 0 to 1. As a preferred approach, the specific data sources and standardization logic configurations for the three indicators are as follows: the probability score is directly extracted from the highest probability value of the target category output by the aforementioned Softmax layer; the severity score is obtained by linearly normalizing the ratio of the current aquifer water pressure mutation value measured by sensors to the historical safety extreme value; and the exposure level score is generated by comprehensively mapping the actual number of workers and the value of fixed assets and equipment in the current area through an interface call to the underground personnel positioning system.

[0071] Based on the comprehensive risk index calculated in real time, a comparator divides the warning standards into four levels according to set thresholds: when the risk index is less than 0.3, a blue warning is issued; when the risk index is greater than or equal to 0.3 and less than 0.6, a yellow warning is issued; when the risk index is greater than or equal to 0.6 and less than 0.85, an orange warning is issued; when the risk index is greater than or equal to 0.85 and the associated safety system detects personnel working in the current area, the highest level red warning is directly triggered. The threshold boundaries of 0.3, 0.6, and 0.85 are statistically calibrated based on the risk distribution range of historical similar water hazard events in this mining area during the offline phase. In practical applications, they can be dynamically configured according to changes in geological conditions.

[0072] S503 generates structured early warning information and activates multi-channel push and control linkage with underlying physical devices. When For example, when an early warning status is triggered, a structured early warning message is automatically generated. This message is configured to include the three-dimensional spatial location of the hazard, the hazard type, risk level, confidence level data, and handling suggestions generated by a cloud-based large model. After receiving the message, the multi-channel push module distributes it in parallel to various terminal interfaces: triggering the audible and visual alarms at the underground site for immediate physical alarm; pushing the information to the mobile devices of management personnel via SMS gateway and mobile application; dispatching the mine's internal broadcast system to issue a voice warning; and simultaneously rendering a graphical highlighting indicator on the monitoring screen of the ground dispatch center.

[0073] As a preferred industrial linkage method, it maintains a bypass listening state under normal monitoring mode. When the current state is determined to be a high-risk state under orange or red alert, the linkage module automatically generates a control protocol frame and sends it to the PLC node, forcibly triggering the start of drainage equipment such as water pumps at designated drainage nodes underground. Simultaneously, it performs shutdown and interlock operations on mining equipment such as coal mining machines and tunneling machines associated with the three-dimensional spatial location of the potential hazard, realizing automated safety isolation from software early warning to physical equipment intervention. Based on this early intervention mechanism, early signs of water hazard incubation can be identified, with an expected early warning time of 24 to 72 hours. Moreover, the prediction time window accuracy for disasters such as water inrush in old workings is expected to be controlled within a fluctuation range of about 4 hours.

[0074] S504, establish a closed-loop mechanism for verifying early warning results and iteratively optimizing the model online.

[0075] To address the issue of scene adaptability degradation in deep learning models over long-term operation, a corresponding early warning result verification pipeline can be deployed. When on-site personnel receive push notifications and proceed to the warning area for verification, they use mobile terminals to feed back the actual physical condition data of the potential hazards to the central server database. This adds a data label with accurate and realistic meaning to the original prediction results, forming a labeled dataset. A dynamic sample library is established, with a fixed update cycle of one week. At the end of this cycle, newly added on-site labeled data is extracted and added to the training set. The error between the model's predicted output and the manually verified labels is calculated using the cross-entropy loss function. A small learning rate fine-tuning strategy is employed to update some network weight parameters of the large vertical model in the cloud and the small model at the edge online, thereby preventing catastrophic forgetting of the model.

[0076] Correspondingly, by comparing the model's predictions with the actual conditions reported by humans, specific weaknesses causing false alarms or missed alarms in the model can be identified. Based on the characteristics of these weaknesses, encrypted data collection commands can be issued to specifically collect sensor data and image samples under such special hydrogeological conditions or extreme environments, allowing the model to perform fine-tuning training. Through this closed-loop feedback pipeline, the model's decision boundaries can be continuously corrected, which is expected to ensure the adaptability and long-term effectiveness of the coal mine disaster prevention system to geological changes.

[0077] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a coal mine water hazard identification method based on size model fusion.

[0078] Specific application examples: To better understand the technical solution of this invention, the working process and application effect of the solution are described in detail below, taking into account the actual operating conditions of the 8202 coal mining face in a certain mining area.

[0079] The 8202 working face has complex hydrogeological conditions, with an Ordovician limestone aquifer on its floor and a concealed fault-conducting zone ahead of the advancing direction. Environmental parameters exhibit typical harsh working conditions: ambient light is consistently below 50 lux, and local dust concentration exceeds 100 mg / m³.

[0080] Data Acquisition and Edge Screening: The intelligent water pressure and level sensor deployed on the 8202 working face detected a slow increase in water pressure of 0.15 MPa over 15 minutes in timed mode, triggering a high-frequency (100Hz) event recording mode. Simultaneously, a high-definition camera captured visual images of subtle changes in the reflectivity of the rock wall under low light and high dust conditions. After preprocessing, an 8.5MB INT8-quantified water hazard prevention model was executed for inference on an edge industrial control computer. This inference took only 85ms (far below the industry standard of 200ms) and identified suspected seepage anomalies.

[0081] Collaborative routing and dynamic fusion: Due to the harsh environment and image fogging, the small model outputs a confidence level of 0.72 for the risk of sudden water inrush in the Ordos Grey River. Based on the routing rules of S401 (confidence level between 0.5 and 0.9), the system determines that the features have boundary ambiguity and immediately reports the time series and image data to the cloud.

[0082] Cloud-based reasoning and knowledge graph enhancement: After receiving data, the cloud-based vertical large model combines extracted 3D geological data (aquifer thickness and fault coordinates) with macroscopic semantic features and local seepage features uploaded from the edge to perform 1536-dimensional splicing and fusion. Simultaneously, the system initiates a RAG search, recalling knowledge graph nodes related to a similar fault-conducting water inrush accident at a mine in 2023. Because the working face is identified as being in a complex geological scenario of fault intersection, the system, based on the S403 rule, increases the voting weight of the large model to 0.7.

[0083] Early Warning Generation and Equipment Linkage: The system ultimately output a comprehensive risk index of 0.78 (orange alert). The system not only concluded there was a risk of water inrush in the Ordovician limestone formation, but also generated an interpretable diagnostic text: the current aquifer pressure exceeds the historical average by 30%, and it is located within the fault's water-conducting zone. Combined with the seepage characteristics of the rock face in the image, and referencing a similar hydrogeological water inrush case at a mine in 2023, there is an extremely high risk of water inrush. The system predicted the water inrush time window of the old workings with an accuracy of ±4 hours and triggered an early warning 68 hours before the disaster occurred. The linkage module immediately issued a command to the underground PLC, automatically starting the drainage pump at the 8202 working face and locking down the tunneling machine, successfully averting a potential water-related accident.

[0084] Experimental verification and effect comparison: To verify the effectiveness of the coal mine water hazard identification system based on large and small model fusion of the present invention, this embodiment conducted a comparative experiment based on the historical dataset of coal mine multi-source water hazard monitoring (CM-MHD) from a large mining area over the past five years. This dataset contains 95% normal state samples and 5% scarce water hazard potential samples. The comparison algorithms include: traditional deep learning visual network (ResNet-50), pure large language multimodal model (baseline large model), and the large and small model fusion architecture of this application.

[0085] The comprehensive recognition performance and timeliness evaluation are shown in Table 1. This invention, through the dynamic fusion of a vertical large model and a small model, fully leverages the local sensitivity of the small model and the global noise resistance of the large model. Compared to a single network, the accuracy of water hazard identification in this invention is improved by 16.5%. Thanks to the semantic reasoning of the large model filtering out false features (such as changes in illumination) and the high-frequency feature capture of the small model, the false alarm rate of this invention is reduced to 4.2% ( 5%), the underreporting rate dropped to 1.5% ( 2%. Regarding timeliness and resource consumption, the edge inference latency of this invention is controlled at 92ms ( ). The cloud-based collaborative response time is 3.2 seconds (within the 2-5 second range). Overall computational resource consumption is reduced by 64% compared to the pure large model solution and by 32% compared to the ResNet-50 solution. The quantized small model's physical footprint is only 7.8MB. (10MB), making it extremely suitable for embedded deployment.

[0086] Table 1. Comparison of Recognition Performance and Resource Consumption of Different Model Architectures

[0087] Robustness assessment under harsh environments and complex data distributions is crucial in downhole operations, which are often characterized by high dust levels and low illumination. This system enhances environmental robustness through preprocessing enhancement and a fusion decision-making mechanism. Experiments show that when the ambient illumination... 50 lux, dust concentration At a concentration of 100 mg / m³, the accuracy of the traditional ResNet-50 method decreased by more than 22%; while the accuracy of the system of this invention decreased by only 3.8% (controlled within 5%). Addressing the bias of prioritizing normal cases over abnormal cases due to the scarcity of flood damage samples, this system utilizes the small-sample learning ability and feature alignment of a large model, maintaining high sensitivity to the less than 5% of potential hazard samples. Its F1-Score for the minority class (old open-hole flooding) reached 0.94.

[0088] The early warning prediction system has achieved a leap from passive response to proactive prevention. Backtesting shows that, for real historical water inrush accidents, this system, through cross-modal fusion tracking of microseismic frequency anomalies and slow water pressure increases, can issue a clear warning on average 48.5 hours before the actual occurrence of the water inrush (within the 24-72 hour range). Furthermore, the prediction error for the water inrush time point is stably controlled within ±4 hours, thus gaining valuable golden rescue time for the evacuation of underground personnel and grouting to plug the water.

[0089] Please see Figure 2 , Figure 2This is a schematic diagram of the structure of a coal mine water hazard identification device based on size model fusion, provided in an embodiment of this application. Figure 2 As shown, the device 200 includes: The first processing module 201 is used to collect multi-source water hazard monitoring data and perform synchronous alignment preprocessing to generate a multi-dimensional feature matrix dataset. The second processing module 202 is used to construct a large vertical model for coal mine water hazard identification based on a multi-dimensional feature matrix dataset, and to introduce the large vertical model for coal mine water hazard identification to train the small water hazard prevention and control model through knowledge distillation. The third processing module 203 is used to input the real-time collected multi-dimensional feature matrix dataset into the water hazard prevention and control small model deployed at the edge for screening, and trigger the cloud-based coal mine water hazard identification vertical large model to perform collaborative calculations based on the target category confidence output by the water hazard prevention and control small model. The fourth processing module 204 is used to extract the global hydrological semantic features output by the vertical large model for coal mine water hazard identification and the local detailed features output by the small model for water hazard prevention and control during collaborative operation, perform feature splicing and fusion, and adjust the voting weights of the vertical large model for coal mine water hazard identification and the small model for water hazard prevention and control in combination with the hydrogeological scene status, and perform joint inference. The fifth processing module 205 is used to assess the risk level based on the joint reasoning results and send disaster prevention linkage control instructions.

[0090] In one implementation, the first processing module 201 can be used to: Acquire time-series data including water pressure, water level, water quality, and microseismic data, as well as visual image data and temperature distribution field matrix data; The time series data were processed by a combination of median filtering and wavelet denoising algorithm. Outlier data points were removed according to the 3σ criterion, and the Min-Max standardization method was used to unify the dimensions of the time series data. A global baseline timestamp sequence is set. For missing data values ​​at non-baseline time points, a linear interpolation algorithm is used to calculate the mapping value at the target time point, and time series data of different frequencies are reconstructed into a time series feature matrix with a unified timestamp. The dark channel prior dehazing algorithm and the Retinex image enhancement algorithm are used to process the visual image data. The processed visual image data, temperature distribution field matrix data and temporal feature matrix are then merged to generate a multi-dimensional feature matrix dataset.

[0091] In one implementation, the second processing module 202 can be used to: Constructing a multimodal feature encoder and generative network for a large-scale vertical model of coal mine water hazard identification; Historical flood damage case records are obtained, and features are extracted from the historical flood damage case records and the multi-dimensional feature matrix dataset through a multimodal feature encoder to generate text feature vectors, image feature vectors and time-series feature vectors. A text-guided dual-branch cross-attention mechanism is adopted, which uses the text feature vector as the query vector and performs dot product calculation and weighted summation with the image feature vector and the temporal feature vector respectively. The output results are then added element-wise to generate global hydrological semantic features. Using global hydrological semantic features as query conditions, a search is performed in a pre-defined coal mine hydrological knowledge graph. The results are generated by autoregressive decoding through a generative network combined with the knowledge of the associated graph obtained from the search, thus completing the construction of a large-scale vertical model for coal mine water hazard identification.

[0092] In one implementation, the second processing module 202 can be used to: The large vertical model for coal mine water hazard identification is used as the teacher model, and the small water hazard prevention and control model is used as the student model. Extract the category probability distribution output by the teacher model as a soft label, and cascade a feature adaptation layer between the corresponding intermediate feature extraction layers of the teacher model and the student model; By combining the soft label distribution divergence loss, the L2 loss between intermediate feature extraction layers, and the cross-entropy loss of the hard labels of real samples, backpropagation is performed to iteratively update the weight parameters of the small water hazard prevention and control model. Network quantization and structured pruning operations were performed on the water hazard prevention and control mini-model that had completed knowledge distillation training.

[0093] In one implementation, the third processing module 203 can be used for: Set a first preset threshold and a second preset threshold, wherein the first preset threshold is greater than the second preset threshold; When the confidence level of the target category is greater than or equal to the first preset threshold, the judgment result of the water hazard prevention and control mini-model is used as the final identification result. When the confidence level of the target category is within the range of the second preset threshold to the first preset threshold, the multi-dimensional feature matrix dataset collected in real time is uploaded to the cloud server, triggering the collaborative operation of the vertical large model for coal mine water hazard identification.

[0094] In one implementation, the fourth processing module 204 can be used for: The global hydrological semantic features and local detailed features are concatenated and cascaded along the channel dimension to form the initial joint features; The initial joint features are input into the channel attention network module, and the gating weight coefficients corresponding to each feature channel are generated through a trainable weight matrix and a non-linear activation function. The initial joint features are element-wise weighted by gating weight coefficients to output a fused feature vector.

[0095] In one implementation, the fourth processing module 204 can be used for: Analyze the current hydrogeological conditions of the working face; When the hydrogeological scene is determined to be in a fault intersection zone, a water-rich aquifer zone, or a fault fracture zone, increase the voting weight of the large vertical model for coal mine water hazard identification and decrease the voting weight of the small water hazard prevention and control model. When the hydrogeological scenario is determined to be a normal single stratum or other conventional hydrogeological scenario, the voting weight of the large vertical model for coal mine water hazard identification should be reduced and the voting weight of the small water hazard prevention and control model should be increased.

[0096] In one implementation, the fourth processing module 204 can also be used for: The multi-class probability set output by the vertical large model for coal mine water hazard identification and the multi-class probability set output by the small model for water hazard prevention and control are horizontally concatenated by feature dimension to construct a secondary meta-feature matrix. Using the secondary meta-feature matrix as input data and the manually verified real water hazard category results as supervision labels, a meta-classifier based on logistic regression algorithm is trained to establish the mapping relationship between the secondary meta-feature matrix and the posterior probability of the final water hazard category, and to fix the decision boundary in the dual-end model fusion state.

[0097] In one implementation, the fifth processing module 205 can be used to: Based on the results of joint reasoning, the highest probability value of the corresponding water hazard category is obtained. A three-dimensional risk assessment index system covering probability, severity and exposure degree is established. The relative weights of probability, severity and exposure degree are calculated by using the analytic hierarchy process. The scores of each standardized index are multiplied by their corresponding relative weights and summed to calculate the comprehensive risk index. The risk level is then assessed based on the set threshold range in which the comprehensive risk index is located. Based on the joint reasoning results and risk level, a structured early warning message is generated, which includes the three-dimensional spatial location of the hazard, the hazard type, the risk level, and the handling suggestions. The structured early warning message is then pushed to the terminal. When the risk level in the structured early warning message reaches the set high-risk warning level, a control protocol frame is sent to the downhole programmable logic controller through the communication interface, which will activate the drainage equipment and lock the mining equipment associated with the three-dimensional spatial location of the hidden danger. The system receives feedback on the actual physical condition of potential hazards after manual review and updates the dynamic sample library. When the set update cycle arrives, it extracts newly labeled data from the dynamic sample library to conduct online iterative fine-tuning training of the large vertical model for coal mine water hazard identification and the small model for water hazard prevention and control.

[0098] It should be noted that the foregoing explanation of the coal mine water hazard identification method based on size model fusion also applies to the coal mine water hazard identification device based on size model fusion in this embodiment, and will not be repeated here.

[0099] To implement the above embodiments, this application also proposes an electronic device. Please see [link to relevant documentation]. Figure 3 , Figure 3 This is a schematic diagram of the structure of the electronic device provided in an embodiment of this application. For example... Figure 3 As shown, the electronic device 300 includes: a processor 301, and a memory 302 communicatively connected to the processor 301; the memory 302 stores computer execution instructions; the processor 301 executes the computer execution instructions stored in the memory to implement the method provided in the foregoing embodiments.

[0100] To implement the above embodiments, this application also proposes a storage medium storing instructions that, when executed on an electronic device, cause the electronic device to perform the methods provided in the foregoing embodiments.

[0101] To implement the above embodiments, this application also proposes a program product, including at least one of a program and instructions, wherein when the program and instructions are executed by an electronic device, they implement the steps of the method provided in the foregoing embodiments.

[0102] It should be noted that the acquisition, transmission, storage, use, and processing of data in this application comply with the relevant provisions of national laws and regulations and do not violate public order and good morals.

[0103] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0104] It is worth noting that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, they do not mean that the applicant has used or necessarily used the solution.

[0105] In the description of this application, unless otherwise stated, " / " means "or", for example, A / B can mean A or B; "and / or" in this document is merely a description of the relationship between related objects, indicating that there can be three relationships, for example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone.

[0106] In the foregoing descriptions of the embodiments, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0107] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0108] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0109] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0110] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0111] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0112] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0113] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for identifying coal mine water hazards based on large-scale model fusion, characterized in that, include: Collect multi-source water hazard monitoring data and perform synchronous alignment preprocessing to generate a multi-dimensional feature matrix dataset; Based on the multi-dimensional feature matrix dataset, a large vertical model for coal mine water hazard identification is constructed, and the large vertical model for coal mine water hazard identification is introduced to train the small water hazard prevention and control model through knowledge distillation. The multi-dimensional feature matrix dataset collected in real time is input into the water hazard prevention and control mini-model deployed at the edge for screening. Based on the target category confidence score output by the water hazard prevention and control mini-model, the coal mine water hazard identification vertical large model in the cloud is triggered to perform collaborative calculation. During the collaborative operation, the global hydrological semantic features output by the vertical large model for coal mine water hazard identification and the local detailed features output by the small model for water hazard prevention and control are extracted, and the features are spliced ​​and fused. The voting weights of the vertical large model for coal mine water hazard identification and the small model for water hazard prevention and control are adjusted in combination with the hydrogeological scene status, and joint inference is performed. The risk level is assessed based on the joint reasoning results, and disaster prevention and control instructions are sent.

2. The method according to claim 1, characterized in that, Multi-source water hazard monitoring data were collected and synchronously aligned preprocessed to generate a multi-dimensional feature matrix dataset, including: Acquire time-series data including water pressure, water level, water quality, and microseismic data, as well as visual image data and temperature distribution field matrix data; The time series data is processed using a median filtering algorithm combined with a wavelet denoising algorithm, and outlier data points are removed according to the 3σ criterion. The Min-Max normalization method is used to unify the dimensions of the time series data. A global reference timestamp sequence is set. For missing data values ​​at non-reference times, a linear interpolation algorithm is used to calculate the mapping value at the target times, and the time series data of different frequencies are reconstructed into a time series feature matrix with a unified timestamp. The visual image data is processed using a dark channel prior dehazing algorithm and a Retinex image enhancement algorithm. The processed visual image data, the temperature distribution field matrix data, and the temporal feature matrix are then merged to generate the multi-dimensional feature matrix dataset.

3. The method according to claim 1, characterized in that, Based on the aforementioned multi-dimensional feature matrix dataset, a large-scale vertical model for coal mine water hazard identification is constructed, including: Construct a multimodal feature encoder and generator network for the vertical large-scale model of coal mine water hazard identification; Historical flood damage case records are obtained, and features are extracted from the historical flood damage case records and the multi-dimensional feature matrix dataset by the multimodal feature encoder to generate text feature vectors, image feature vectors and time-series feature vectors. A text-guided dual-branch cross-attention mechanism is adopted, in which the text feature vector is used as the query vector, and dot product calculation and weighted summation are performed with the image feature vector and the temporal feature vector respectively. The output results are then added element-wise to generate the global hydrological semantic features. Using the global hydrological semantic features as query conditions, a search is performed in the preset coal mine hydrological knowledge graph. The generative network, combined with the retrieved associated graph knowledge, performs autoregressive decoding to generate inference results, thus completing the construction of the vertical large model for coal mine water hazard identification.

4. The method according to claim 1, characterized in that, The large-scale vertical model for coal mine water hazard identification is introduced to train the small-scale water hazard prevention and control model through knowledge distillation, including: The large vertical model for coal mine water hazard identification is used as the teacher model, and the small water hazard prevention and control model is used as the student model. The category probability distribution output by the teacher model is extracted as a soft label, and a feature adaptation layer is cascaded between the corresponding intermediate feature extraction layers of the teacher model and the student model. The weight parameters of the water hazard prevention and control mini-model are updated iteratively by combining the soft label distribution divergence loss, the L2 loss between the intermediate feature extraction layers, and the cross-entropy loss of the hard labels of the real samples; The water hazard prevention and control mini-model that has completed knowledge distillation training is subjected to network quantization and structured pruning operations.

5. The method according to claim 1, characterized in that, Based on the target category confidence score output by the small-scale water hazard prevention and control model, the large-scale vertical model for coal mine water hazard identification in the cloud is triggered to perform collaborative computation, including: Set a first preset threshold and a second preset threshold, wherein the first preset threshold is greater than the second preset threshold; When the confidence level of the target category is greater than or equal to the first preset threshold, the judgment result of the water hazard prevention and control mini-model is used as the final identification result. When the confidence level of the target category is within the range of the second preset threshold to the first preset threshold, the multi-dimensional feature matrix dataset collected in real time is uploaded to the cloud server, triggering the collaborative operation of the coal mine water hazard identification vertical large model.

6. The method according to claim 1, characterized in that, The global hydrological semantic features output by the large-scale vertical model for coal mine water hazard identification are extracted and the local detailed features output by the small-scale water hazard prevention model are concatenated and fused, including: The global hydrological semantic features and the local detail features are concatenated and cascaded along the channel dimension to form an initial joint feature; The initial joint features are input into the channel attention network module, and the gating weight coefficients corresponding to each feature channel are generated through a trainable weight matrix and a nonlinear activation function. The initial joint features are element-wise weighted and mapped using the gating weight coefficients to output a fused feature vector.

7. The method according to claim 1, characterized in that, The voting weights of the large-scale vertical model for coal mine water hazard identification and the small-scale water hazard prevention model are adjusted based on the hydrogeological scene conditions, including: Analyze the current hydrogeological scene status of the working face; When the hydrogeological scene is determined to be in a fault intersection zone, a water-rich aquifer zone, or a fault fracture zone, the voting weight of the large vertical model for coal mine water hazard identification is increased and the voting weight of the small water hazard prevention and control model is decreased. When the hydrogeological scene is determined to be a normal single stratum or other conventional hydrogeological scene, the voting weight of the coal mine water hazard identification vertical large model is reduced and the voting weight of the water hazard prevention and control small model is increased.

8. The method according to claim 1, characterized in that, After performing the joint inference, the following is also included: The multi-class probability set output by the vertical large model for coal mine water hazard identification and the multi-class probability set output by the small model for water hazard prevention and control are horizontally concatenated by feature dimension to construct a secondary meta-feature matrix. Using the secondary meta-feature matrix as input data and the manually verified real water hazard category results as supervision labels, a meta-classifier based on logistic regression algorithm is trained to establish the mapping relationship between the secondary meta-feature matrix and the posterior probability of the final water hazard category, thus fixing the decision boundary in the dual-end model fusion state.

9. The method according to claim 1, characterized in that, The risk level is assessed based on the joint reasoning results, and disaster prevention and control instructions are sent, including: Based on the joint reasoning results, the highest probability value of the corresponding water hazard category is obtained, a three-dimensional risk assessment index system covering probability, severity and exposure degree is established, the relative weights of probability, severity and exposure degree are calculated using the analytic hierarchy process, the scores of each standardized index are multiplied by the corresponding relative weights and summed to calculate the comprehensive risk index, and the graded risk level is evaluated based on the set threshold range in which the comprehensive risk index is located. Based on the joint reasoning results and the risk level, a structured early warning message is generated, which includes the three-dimensional spatial location of the hidden danger, the type of hidden danger, the risk level and the handling suggestions, and the structured early warning message is pushed to the terminal. When the risk level in the structured early warning message reaches the set high-risk warning level, a control protocol frame is sent to the downhole programmable logic controller through the communication interface to activate the drainage equipment and lock the mining equipment associated with the three-dimensional spatial location of the hidden danger. The system receives feedback on the actual physical condition of potential hazards after manual review and updates the dynamic sample library. When the set update cycle arrives, it extracts newly labeled data from the dynamic sample library to perform online iterative fine-tuning training on the large vertical model for coal mine water hazard identification and the small model for water hazard prevention and control.

10. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 9.

Images