Continuous monitoring method for composite roof collaborative filling surface

Abstract

The invention discloses a continuous monitoring method for a composite roof collaborative filling surface, and belongs to the field of filling surface monitoring, and the method comprises the steps: arranging a multi-source sensor based on stope and roadway geometry and filling technology, obtaining and carrying out the time-space alignment of a multi-source signal, extracting the space and time features through a graph neural network and a time sequence neural network, and carrying out the time-space alignment of the multi-source signal; a continuous state field is generated through multi-modal attention fusion; the strategy network outputs filling parameter instructions according to a state field and physical constraints to achieve closed-loop regulation and control, and meanwhile, continuous, high-precision and dynamic monitoring and self-adaptive control of the composite roof collaborative filling surface are achieved in combination with an online incremental learning dynamic optimization model.

Description

Technical Field

[0001] This invention belongs to the field of filling surface monitoring, and particularly relates to a continuous monitoring method for composite top co-filling surfaces. Background Technology

[0002] In mining operations, the composite roof and co-filling face is a critical component for ensuring roof stability and operational safety in the goaf. Existing filling monitoring methods largely rely on single-type sensors, such as stress gauges or displacement gauges, which can only acquire localized information and are insufficient to reflect the overall condition of the entire filling face and roadway. Furthermore, traditional methods typically employ periodic manual observation or low-frequency sampling, failing to achieve continuous monitoring of the filling process and thus hindering real-time control and safety early warning capabilities. In addition, existing methods lack comprehensive utilization of multimodal information, making it difficult to model the nonlinear evolution of complex geometries and dynamic filling processes. This results in delayed adjustments to filling parameters, making it impossible to accurately guarantee filling quality and roof stability. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a method for continuous monitoring of composite top co-filling surfaces, comprising:

[0004] Based on the geometry and filling technology of the stope and roadway, multiple raw signals are obtained;

[0005] Based on the deployment parameters and calibration curves of the multi-source raw signals, a spatiotemporally aligned sampling stream is obtained;

[0006] A graph neural network and a temporal neural network are constructed, and the spatiotemporally aligned sampling stream is processed based on the graph neural network and the temporal neural network. A continuous state field is obtained through multimodal attention fusion.

[0007] Based on the continuous state field and physical constraints, the filling parameter command is output based on the policy network to control the filling process in a closed loop.

[0008] Online incremental learning is performed based on closed-loop feedback to complete continuous monitoring and control of the composite top co-filling surface.

[0009] Optionally, the process of acquiring multi-source raw signals based on the geometry and filling technology of the stope and roadway includes:

[0010] Based on the three-dimensional geometric parameters of the mining area and the roadway layout plan, the placement positions of tension sensors, stress gauges, and displacement gauges are determined to obtain stress and displacement signals at the contact surface between the filling material and the roof.

[0011] Based on the slurry transport path and grouting rate in the filling process, acoustic emission sensors and flow meters are deployed to acquire acoustic emission signals and slurry flow parameters during the filling process.

[0012] The video or point cloud data of the mining section and filling surface are acquired through a visual acquisition device to form a multi-source raw signal set for subsequent preprocessing.

[0013] Optionally, the process of obtaining a spatiotemporally aligned sampling stream based on the deployment parameters and calibration curves of the multi-source raw signals includes:

[0014] Based on the deployment parameters of the tension sensor, stress gauge and displacement gauge, the signals of each sensor are synchronized according to a unified timestamp;

[0015] Based on the calibration curves of the acoustic emission sensor and the flow meter, the amplitude and phase of the acquired signal are corrected.

[0016] Then, based on the point cloud data generated by the visual acquisition device and the geometric model of the mining site, a spatial coordinate mapping relationship is established, and various signals are mapped to the same three-dimensional coordinate system;

[0017] Then, through interpolation operations and temporal resampling, a spatiotemporally aligned continuous sampling stream is formed.

[0018] Optionally, the process of constructing the graph neural network and the temporal neural network includes:

[0019] Based on the geometry of the mining area and roadway, the topological relationship between nodes and edges is established, the multi-source signals collected by the sensors are mapped to the corresponding node features, and the spatial correlation is aggregated through graph convolutional layers.

[0020] Based on the temporal sampling stream, the features of each node are input into the temporal neural network, and continuous time segments are recursively updated. Long short-term memory units or gated recurrent units are used to capture dynamic evolution patterns, and joint modeling of spatial topological features and time series features is completed.

[0021] Optionally, the process of processing the spatiotemporally aligned sample stream based on the graph neural network and the temporal neural network, and obtaining a continuous state field through multimodal attention fusion, includes:

[0022] The spatial dependence features of the multi-source spatiotemporal signals collected by the sensors are extracted by a graph neural network, and the temporal dynamic features are extracted by a temporal neural network.

[0023] Different modal features, such as visual data, stress data, and acoustic emission data, are input into the multimodal attention module to calculate the weight distribution of each modality in different spatiotemporal segments.

[0024] Based on the weight distribution, the features of different modalities are weighted and fused to obtain a unified high-dimensional feature representation;

[0025] The high-dimensional features are restored to a spatially continuously distributed state field by a decoder.

[0026] Optionally, the process of inputting different modal features such as visual data, stress data, and acoustic emission data into the multimodal attention module and calculating the weight distribution of each modality in different spatiotemporal segments includes:

[0027] The different modal features are normalized and dimensionally aligned; through the mapping of key vectors, query vectors and numerical vectors, each modal feature is mapped to a unified representation space;

[0028] The attention weight coefficient is calculated based on the similarity between the query vector and the key vector.

[0029] The weighting coefficients are assigned to each spatiotemporal segment to reflect the importance of each modal feature;

[0030] The numerical vectors of each modality are weighted and summed according to the weight coefficients to generate a fusion feature representation under the corresponding spatiotemporal segment.

[0031] Optionally, the process of outputting filling parameter commands based on the policy network to control the filling process in a closed loop according to the continuous state field and physical constraints includes:

[0032] The continuous state field is input into the policy network as the environment state.

[0033] Based on the physical constraints, a constraint function is constructed to limit the strategy search space;

[0034] A set of candidate filling parameters is generated by forward reasoning about the environment state through a policy network.

[0035] The set of candidate filling parameters is filtered by applying constraint functions to obtain feasible solutions that satisfy the constraints;

[0036] The feasible code is decoded into filling parameter instructions; and the filling parameter instructions are fed back to the filling equipment.

[0037] Optionally, the process of performing online incremental learning based on closed-loop feedback to continuously monitor and control the composite top co-filling surface includes:

[0038] The actual working condition data after the filling equipment is executed is compared with the continuous state field to calculate the monitoring error and deviation.

[0039] Adjusting the weight parameters of the strategy network and neural network model based on errors and biases enables online incremental training;

[0040] The updated model is reused for feature extraction, state field prediction, and filling parameter instruction generation;

[0041] The incremental learning and closed-loop control process is executed cyclically, enabling the system to adaptively optimize the filling process during continuous monitoring, thereby achieving dynamic control and stable monitoring of the composite top co-filling surface.

[0042] On the other hand, the present invention also provides an electronic device including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0043] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0044] Compared with the prior art, the present invention has the following advantages and technical effects:

[0045] This invention achieves comprehensive monitoring of the composite roof co-filling surface by deploying multiple types of sensors and acquiring multi-source raw signals based on the geometric features of the stope and roadway and the filling process parameters. The multi-source signals are spatiotemporally aligned using sensor deployment parameters and calibration curves to form a continuous sampling stream, providing a precise data foundation for subsequent feature extraction. A graph neural network is used to model spatial topological dependencies, combined with a temporal neural network to capture dynamic evolution patterns. A multimodal attention mechanism is used to fuse different modal features such as visual, stress, and acoustic emission, enabling continuous state field prediction of the roof and filling body states. Based on this, a strategy network generates filling parameter instructions according to the continuous state field and physical constraints, achieving closed-loop control of the filling process. Simultaneously, the system performs online incremental learning based on closed-loop feedback, dynamically updating the neural network model and strategy network parameters, enabling continuous adaptive optimization of the monitoring system during the filling process. Through the above technical solution, high-precision, continuous, and dynamic monitoring and control of the composite roof co-filling surface can be achieved, providing reliable data support and decision-making basis for filling process optimization and roof stability assurance. Attached Figure Description

[0046] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0047] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0048] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0049] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0050] Example 1

[0051] like Figure 1 As shown, this embodiment provides a method for continuous monitoring of composite top co-filling surfaces, including:

[0052] Based on the geometry and filling technology of the stope and roadway, multiple raw signals are obtained;

[0053] Based on the deployment parameters and calibration curves of the multi-source raw signals, a spatiotemporally aligned sampling stream is obtained;

[0054] A graph neural network and a temporal neural network are constructed, and the spatiotemporally aligned sampling stream is processed based on the graph neural network and the temporal neural network. A continuous state field is obtained through multimodal attention fusion.

[0055] Based on the continuous state field and physical constraints, the filling parameter command is output based on the policy network to control the filling process in a closed loop.

[0056] Online incremental learning is performed based on closed-loop feedback to complete continuous monitoring and control of the composite top co-filling surface.

[0057] Furthermore, the process of acquiring multi-source raw signals based on the geometry and filling technology of the stope and roadway includes:

[0058] Based on the three-dimensional geometric parameters of the mining area and the roadway layout plan, the placement positions of tension sensors, stress gauges, and displacement gauges are determined to obtain stress and displacement signals at the contact surface between the filling material and the roof.

[0059] Based on the slurry transport path and grouting rate in the filling process, acoustic emission sensors and flow meters are deployed to acquire acoustic emission signals and slurry flow parameters during the filling process.

[0060] The video or point cloud data of the mining section and filling surface are acquired through a visual acquisition device to form a multi-source raw signal set for subsequent preprocessing.

[0061] In continuous monitoring methods for composite roof and backfill faces, the process of acquiring multi-source raw signals based on the geometry and backfilling process of the stope and roadway is the data foundation and key link of the entire monitoring system. Its main purpose is to comprehensively and accurately capture multi-dimensional information about the roof, backfill body, and backfilling process in complex mining environments, providing reliable raw data support for subsequent data preprocessing, feature extraction, and neural network modeling. This process includes three main steps: sensor deployment, signal acquisition, and multi-source data acquisition and integration.

[0062] Based on the three-dimensional geometric parameters of the mining area and the roadway layout plan, the locations of tension sensors, stress gauges, and displacement gauges were determined. Through mechanical characteristic analysis of the contact surface between the roof and the backfill, stress concentration areas and key support nodes were selected as sensor placement points to ensure real-time reflection of the stress state of the roof and the deformation behavior of the backfill. Tension sensors are used to detect the tensile or compressive state between the roof and the support structure, stress gauges are used to record local stress distribution, and displacement gauges are used to measure the displacement changes of the roof and the backfill. After sensor placement, preliminary calibration was performed to obtain zero-point and sensitivity parameters, ensuring data acquisition accuracy.

[0063] Based on the grout delivery path and injection rate in the filling process, acoustic emission sensors and flow meters are deployed to acquire dynamic information during the filling process. Acoustic emission sensors capture high-frequency signals generated by the propagation of microcracks within the filling material, grout flow impact, and local instability of the filling surface, reflecting the microscopic dynamic changes during the filling process in real time. Flow meters record the grout injection rate and delivery parameters, providing a basis for analyzing filling uniformity, density, and flow regime. By combining the injection path, injection speed, and sensor locations, a multi-source signal that synchronously reflects the filling dynamics and structural response is formed.

[0064] Images or point cloud data of the stope cross-section and filling surface are acquired through visual acquisition devices to achieve digital representation of spatial geometric information. High-resolution cameras or 3D laser scanning devices are used to acquire images of the stope cross-section frame by frame. A continuous 3D surface model is generated through point cloud stitching and spatial filtering to reflect roof settlement, fracture propagation, and the surface condition of the filling body. Video data can be used to dynamically monitor abnormalities during the filling process, such as uneven grout flow or localized roof deformation. All multi-source signals are recorded and initially processed using unified timestamps and spatial coordinates to form a multi-source raw signal set for subsequent data preprocessing.

[0065] Furthermore, the process of obtaining a spatiotemporally aligned sampling stream based on the deployment parameters and calibration curves of the multi-source raw signals includes:

[0066] Based on the deployment parameters of the tension sensor, stress gauge and displacement gauge, the signals of each sensor are synchronized according to a unified timestamp;

[0067] Based on the calibration curves of the acoustic emission sensor and the flow meter, the amplitude and phase of the acquired signal are corrected.

[0068] Then, based on the point cloud data generated by the visual acquisition device and the geometric model of the mining site, a spatial coordinate mapping relationship is established, and various signals are mapped to the same three-dimensional coordinate system;

[0069] Then, through interpolation operations and temporal resampling, a spatiotemporally aligned continuous sampling stream is formed.

[0070] In the continuous monitoring method for composite top co-filling surfaces, obtaining a spatiotemporally aligned sampling stream based on the deployment parameters and calibration curves of multi-source raw signals is a key step for subsequent feature extraction and neural network analysis. Its main objective is to unify data from different sensors and modalities in time and space, forming a continuously processable multi-source data stream, thereby ensuring monitoring accuracy and modeling reliability. This process includes four core components: time synchronization, amplitude and phase correction, spatial mapping, and interpolation and resampling.

[0071] First, based on the deployment parameters of the tension sensor, stress gauge, and displacement gauge, the signals from each sensor are synchronized according to a unified timestamp. Let the sampling time series of each sensor be t. i and measured value x i (t i Map it to a unified time axis T = {t1, t2, ..., t}. n On}, it can be represented as:

[0072]

[0073] This step ensures that the signals from different sensors correspond precisely in the time dimension, laying the foundation for multimodal feature fusion.

[0074] Secondly, based on the calibration curves of the acoustic emission sensor and the flow meter, amplitude and phase corrections are performed on the acquired signal. The calibration function f is obtained through calibration experiments. i (·) and phase correction function φ i (·), which transforms the original signal y i (t) is converted to a calibrated signal:

[0075]

[0076] This step eliminates sensor nonlinearity errors and phase drift, ensuring that the amplitude and phase of the acoustic emission and flow signals are consistent with the actual physical quantities.

[0077] Third, a spatial coordinate mapping relationship is established between the point cloud data generated by the visual acquisition device and the geometric model of the acquisition site, mapping various signals to the same three-dimensional coordinate system. Let the spatial coordinates of the point cloud be P. v ={(xk ,y k ,z k The coordinates of the mining area geometric model are P. g Coordinate mapping is achieved through a rigid transformation matrix R and a translation vector t:

[0078]

[0079] This mapping ensures that the data from each sensor have a consistent spatial reference and can be precisely aligned with the geometry of the top plate and the filling body.

[0080] Finally, through interpolation and temporal resampling, a spatiotemporally aligned continuous sampling stream is formed. Let the original sampling points be a discrete sequence {(t... i ,x i The data is then resampled to a unified time grid T using the interpolation function I(·).

[0081] Simultaneously, the spatial features are continuously gridded to form a unified three-dimensional spatiotemporal continuous sampling stream S(t,x,y,z), which can be used for feature extraction and fusion analysis of subsequent graph neural networks and temporal neural networks.

[0082] Furthermore, the process of constructing the graph neural network and the temporal neural network includes:

[0083] Based on the geometry of the mining area and roadway, the topological relationship between nodes and edges is established, the multi-source signals collected by the sensors are mapped to the corresponding node features, and the spatial correlation is aggregated through graph convolutional layers.

[0084] Based on the temporal sampling stream, the features of each node are input into the temporal neural network, and continuous time segments are recursively updated. Long short-term memory units or gated recurrent units are used to capture dynamic evolution patterns, and joint modeling of spatial topological features and time series features is completed.

[0085] In the continuous monitoring method for composite top-coordinated filling surfaces, constructing a graph neural network (GNN) and a recurrent neural network (RNN) is a key step in achieving joint modeling of spatial topology and temporal dynamic features, providing high-dimensional feature representations for filling surface state prediction and closed-loop control. This process mainly includes two core components: spatial topological feature modeling and temporal dynamic feature modeling.

[0086] First, the topological relationships between nodes and edges are established based on the geometry of the mining area and roadways. Key locations on the roof, measuring points of the backfill, and important nodes in the roadways are designated as nodes v in the graph structure. i ∈V, node features are multi-source signals x collected by sensors. i(t) constitutes the physical proximity or mechanical coupling relationship between nodes, defined as edge e. ij ∈E, forming a graph G=(V,E). The graph convolutional layer extracts spatial correlation information by aggregating the features of neighboring nodes. Its basic calculation formula is:

[0087]

[0088] in, Let i be the feature vector of node i in the l-th layer. Let c be the set of neighbors of node i. ij W is the normalization coefficient. (l) Here, σ is the trainable weight matrix, and σ(·) is the activation function. Through multi-layer graph convolution, local topological relationships and spatial dependencies can be captured, thereby encoding complex mining geometry and tunnel structure information into node representations.

[0089] Secondly, based on the temporal sampling stream, the node features after graph convolution are input into the temporal neural network to recursively update continuous time segments in order to capture the dynamic evolution of the filling process.

[0090] Let h be the characteristic of node i at time t. i (t), using Long Short-Term Memory (LSTM) or Gated Recurrent Unit (GRU) to process the time series, a typical LSTM update formula is:

[0091]

[0092] Among them, f t i t ,o t These are the forget gate, input gate, and output gate, respectively. ⊙ represents element-wise multiplication. C t is the memory unit, and W and b are the network weights and biases. Through recursive iteration, the changing trends of node features over time can be captured, and historical information can be preserved.

[0093] Finally, the spatial features obtained from the graph convolutional layer are jointly modeled with the time-series features output by the RNN to form a comprehensive high-dimensional feature vector z. i (t), used for subsequent multimodal fusion and state field prediction. Feature fusion can be expressed as:

[0094]

[0095] Here, φ(·) is the feature fusion function, which can be a simple concatenation, weighted fusion, or attention mechanism. This joint modeling process ensures a complete representation of the spatial topological dependence and temporal dynamic changes of the filling surface, providing reliable input features for closed-loop control and continuous monitoring.

[0096] Furthermore, the process of processing the spatiotemporally aligned sampling stream based on the graph neural network and the temporal neural network, and obtaining a continuous state field through multimodal attention fusion, includes:

[0097] The spatial dependence features of the multi-source spatiotemporal signals collected by the sensors are extracted by a graph neural network, and the temporal dynamic features are extracted by a temporal neural network.

[0098] Different modal features, such as visual data, stress data, and acoustic emission data, are input into the multimodal attention module to calculate the weight distribution of each modality in different spatiotemporal segments.

[0099] Based on the weight distribution, the features of different modalities are weighted and fused to obtain a unified high-dimensional feature representation;

[0100] The high-dimensional features are restored to a spatially continuously distributed state field by a decoder.

[0101] Multi-source spatiotemporal signals collected by sensors are input into a graph neural network to extract spatially dependent features. By constructing a geometric topology map of the stope and roadways, key locations on the roof, measuring points of the backfill, and important nodes in the roadways are used as graph structure nodes. Node features are composed of sensor-collected numerical values, and the physical proximity or mechanical coupling relationships between nodes are defined as graph edges. Graph convolutional layers effectively capture spatial correlations and local topological features by aggregating neighbor node features, providing a high-dimensional spatial representation for subsequent time series modeling. Through multi-layer graph convolution, the complex geometric structure of the stope and the layout information of the roadways can be encoded, achieving a comprehensive perception of the spatial state of the backfill surface.

[0102] The node spatial features extracted by graph convolution are input into a temporal neural network to recursively update continuous time segments, capturing the dynamic evolution of the infill process. Employing Long Short-Term Memory (LSTM) units or Gated Recurrent Units (GRUs), the system retains historical information, modeling the changing states of the roof and infill over time, thus reflecting the stress and deformation dynamics of the infill surface at different stages. This step ensures continuity in the temporal dimension and the integrity of the dynamic features.

[0103] Spatial-temporal features from different modalities, such as visual data, stress data, and acoustic emission data, are input into a multimodal attention module. The importance of each modality in different spatiotemporal segments is evaluated, and a weight distribution is calculated. Subsequently, based on the weight distribution, the features of each modality are weighted and fused to generate a unified high-dimensional feature representation. Through this mechanism, the system can integrate information from different types of sensors, making state prediction more comprehensive and accurate, while improving the accuracy of anomaly detection and monitoring.

[0104] The fused high-dimensional features are input into the decoder, which restores the features to a spatially continuous state field. The decoder maps the node-level high-dimensional features to a three-dimensional spatial grid, enabling a continuous representation of the stress, deformation, and state of the entire filling surface top plate. The generated continuous state field can be used for closed-loop control, filling parameter optimization, and anomaly identification.

[0105] Furthermore, the process of inputting different modal features such as visual data, stress data, and acoustic emission data into the multimodal attention module and calculating the weight distribution of each modality in different spatiotemporal segments includes:

[0106] The different modal features are normalized and dimensionally aligned; through the mapping of key vectors, query vectors and numerical vectors, each modal feature is mapped to a unified representation space;

[0107] The attention weight coefficient is calculated based on the similarity between the query vector and the key vector.

[0108] The weighting coefficients are assigned to each spatiotemporal segment to reflect the importance of each modal feature;

[0109] The numerical vectors of each modality are weighted and summed according to the weight coefficients to generate a fusion feature representation under the corresponding spatiotemporal segment.

[0110] Normalization and dimension alignment are performed on different modal features. The raw features of visual, stress, and acoustic emission data differ in dimensions, scale, and resolution. Normalization maps each modal feature to a unified numerical range, while dimension alignment adjusts the vector dimensions of each modal feature, enabling computation within the same feature space. This step ensures the comparability and feasibility of feature fusion across different sensor types, laying the foundation for subsequent attention calculations.

[0111] By mapping key vectors, query vectors, and numerical vectors, the normalized modal features are projected onto a unified representation space. This mapping encodes the original features into high-dimensional vector forms more suitable for attention mechanisms, enabling different modalities to interact within the same representation space. Through this step, signal features from different sensors can undergo similarity calculation and information fusion within the same semantic space.

[0112] Attention weight coefficients are calculated based on the similarity between the query vector and the key vector. This step assesses the importance of each modality in the current time segment and spatial location, assigning a corresponding level of attention to each modality. The magnitude of the weight reflects the contribution of the modality's features to the fusion result, enabling the system to adaptively adjust its dependence on modal information in different scenarios.

[0113] For each spatiotemporal segment, attention weight coefficients are assigned to the corresponding feature vectors. By finely allocating weights in the temporal and spatial dimensions, the system can dynamically reflect the importance of each modal feature at different locations and time periods. For example, within a certain time segment, visual features may be more reflective of the evolution of roof cracks, while acoustic emission data may be more sensitive to slurry flow anomalies. Weight allocation ensures that the fusion process is dynamically adaptable in spatiotemporal dimensions.

[0114] The numerical vectors of each modality are weighted and summed according to the assigned weight coefficients to generate a fusion feature representation for the corresponding spatiotemporal segment. Through weighted summation, the system integrates the advantageous features of each modality to form a unified and high-dimensional feature representation, which not only preserves the integrity of multimodal information but also highlights the features that contribute most to the current monitoring target.

[0115] Furthermore, the process of outputting filling parameter commands based on the policy network to control the filling process in a closed loop according to the continuous state field and physical constraints includes:

[0116] The continuous state field is input into the policy network as the environment state.

[0117] Based on the physical constraints, a constraint function is constructed to limit the strategy search space;

[0118] A set of candidate filling parameters is generated by forward reasoning about the environment state through a policy network.

[0119] The set of candidate filling parameters is filtered by applying constraint functions to obtain feasible solutions that satisfy the constraints;

[0120] The feasible code is decoded into filling parameter instructions; and the filling parameter instructions are fed back to the filling equipment.

[0121] A continuous state field is input as the environmental state into the policy network. This continuous state field, obtained through the aforementioned multimodal attention fusion and neural network processing, reflects the spatial-temporal dynamic distribution of the roof, infill body, and infill surface, including stress, displacement, cracks, and flow regime information. Using this continuous state field as input environmental state, the policy network encodes the overall and local state information of the infill surface into a vector representation that the network can process, providing real-time environmental awareness for subsequent parameter generation.

[0122] Constraint functions are constructed based on physical constraints to limit the strategy search space. These physical constraints include filling strength requirements, slurry flow rate limits, roof bearing capacity, and roadway geometric boundary conditions. By encoding these constraints into constraint functions, the strategy network can automatically eliminate schemes that do not meet safety or technological requirements when generating candidate filling parameters, thereby ensuring that the generated filling parameters are feasible in actual operation.

[0123] A policy network performs forward reasoning on the environmental state to generate a set of candidate filling parameters. Based on the input continuous state field vector, the policy network performs calculations through a multi-layer neural network to output a series of possible combinations of filling parameters, including grout injection rate, pump pressure, grout concentration, and grouting path. These candidate parameter sets represent potential operational schemes that can be executed under the current filling surface state.

[0124] The candidate filling parameter set is filtered using constraint functions to obtain feasible solutions that satisfy physical and technological constraints. By verifying the constraints of each candidate parameter combination, the system eliminates parameters that do not meet the requirements of roof safety, filling uniformity, or equipment operating limits, ensuring that the final feasible solution meets the technological requirements without causing risks such as roof instability or equipment overload.

[0125] The selected feasible decoders are converted into filling parameter commands. The decoding process transforms the vector-form parameters output by the network into specific equipment control commands, such as pumping rate commands, grouting pressure adjustment commands, and grouting valve opening and closing commands, so that they can directly drive the filling equipment to perform operations.

[0126] The generated filling parameter commands are fed back to the filling equipment in real time, achieving closed-loop control. The filling equipment adjusts the slurry delivery rate, pump pressure, and grouting path according to the commands, thereby dynamically regulating the filling process. At the same time, newly collected status information during the filling process will be re-entered into the strategy network to achieve real-time updates of the environmental status and continuous closed-loop control, ensuring that the composite top collaborative filling surface remains stable and safe throughout the entire filling process.

[0127] Furthermore, the process of continuously monitoring and controlling the composite top co-filling surface by performing online incremental learning based on closed-loop feedback includes:

[0128] The actual working condition data after the filling equipment is executed is compared with the continuous state field to calculate the monitoring error and deviation.

[0129] Adjusting the weight parameters of the strategy network and neural network model based on errors and biases enables online incremental training;

[0130] The updated model is reused for feature extraction, state field prediction, and filling parameter instruction generation;

[0131] The incremental learning and closed-loop control process is executed cyclically, enabling the system to adaptively optimize the filling process during continuous monitoring, thereby achieving dynamic control and stable monitoring of the composite top co-filling surface.

[0132] Finally, to reduce the redundancy of high-dimensional features and improve training efficiency, dimensionality reduction optimization is necessary. Principal component analysis (PCA) or an autoencoder can be used to implement nonlinear mapping and obtain a low-dimensional latent representation z;

[0133] z = f θ (X),X≈g φ (z);

[0134] Among them, f θ and g φ These are the encoder and decoder functions, respectively. The feature representation in this low-dimensional latent space not only preserves the main structural information of the monitoring data but also enhances the generalization ability of the neural network model under complex conditions.

[0135] Through the aforementioned feature extraction and representation learning process, the raw multi-source sensor data is effectively transformed into structured high-dimensional or low-dimensional feature representations, providing a solid data foundation for subsequent anomaly identification, state prediction, and intelligent control. The rational design of this process ensures the stability and accuracy of the continuous monitoring system for the composite top co-filling face in complex mining environments.

[0136] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor implements the method when executing the computing program.

[0137] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method.

[0138] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for continuous monitoring of composite top co-filling surfaces, characterized in that, include: Based on the geometry and filling technology of the stope and roadway, multiple raw signals are obtained; Based on the deployment parameters and calibration curves of the multi-source raw signals, a spatiotemporally aligned sampling stream is obtained; A graph neural network and a temporal neural network are constructed, and the spatiotemporally aligned sampling stream is processed based on the graph neural network and the temporal neural network. A continuous state field is obtained through multimodal attention fusion. Based on the continuous state field and physical constraints, the filling parameter command is output based on the policy network to control the filling process in a closed loop. Online incremental learning is performed based on closed-loop feedback to complete continuous monitoring and control of the composite top co-filling surface.

2. The method according to claim 1, characterized in that, The process of acquiring multi-source raw signals based on the geometry and filling technology of the stope and roadway includes: Based on the three-dimensional geometric parameters of the mining area and the roadway layout plan, the placement positions of tension sensors, stress gauges, and displacement gauges are determined to obtain stress and displacement signals at the contact surface between the filling material and the roof. Based on the slurry transport path and grouting rate in the filling process, acoustic emission sensors and flow meters are deployed to acquire acoustic emission signals and slurry flow parameters during the filling process. The video or point cloud data of the mining section and filling surface are acquired through a visual acquisition device to form a multi-source raw signal set for subsequent preprocessing.

3. The method according to claim 1, characterized in that, The process of obtaining a spatiotemporally aligned sampling stream based on the deployment parameters and calibration curves of multi-source raw signals includes: Based on the deployment parameters of the tension sensor, stress gauge and displacement gauge, the signals of each sensor are synchronized according to a unified timestamp; Based on the calibration curves of the acoustic emission sensor and the flow meter, the amplitude and phase of the acquired signal are corrected. Then, based on the point cloud data generated by the visual acquisition device and the geometric model of the mining site, a spatial coordinate mapping relationship is established, and various signals are mapped to the same three-dimensional coordinate system; Then, through interpolation operations and temporal resampling, a spatiotemporally aligned continuous sampling stream is formed.

4. The method according to claim 1, characterized in that, The process of constructing the graph neural network and the temporal neural network includes: Based on the geometry of the mining area and roadway, the topological relationship between nodes and edges is established, the multi-source signals collected by the sensors are mapped to the corresponding node features, and the spatial correlation is aggregated through graph convolutional layers. Based on the temporal sampling stream, the features of each node are input into the temporal neural network, and continuous time segments are recursively updated. Long short-term memory units or gated recurrent units are used to capture dynamic evolution patterns, and joint modeling of spatial topological features and time series features is completed.

5. The method according to claim 1, characterized in that, The process of processing the spatiotemporally aligned sample stream based on the graph neural network and the temporal neural network, and obtaining a continuous state field through multimodal attention fusion, includes: The spatial dependence features of the multi-source spatiotemporal signals collected by the sensors are extracted by a graph neural network, and the temporal dynamic features are extracted by a temporal neural network. Different modal features, such as visual data, stress data, and acoustic emission data, are input into the multimodal attention module to calculate the weight distribution of each modality in different spatiotemporal segments. Based on the weight distribution, the features of different modalities are weighted and fused to obtain a unified high-dimensional feature representation; The high-dimensional features are restored to a spatially continuously distributed state field by a decoder.

6. The method according to claim 5, characterized in that, The process of inputting different modal features such as visual data, stress data, and acoustic emission data into the multimodal attention module and calculating the weight distribution of each modality under different spatiotemporal segments includes: The different modal features are normalized and dimensionally aligned; through the mapping of key vectors, query vectors and numerical vectors, each modal feature is mapped to a unified representation space; The attention weight coefficient is calculated based on the similarity between the query vector and the key vector. The weighting coefficients are assigned to each spatiotemporal segment to reflect the importance of each modal feature; The numerical vectors of each modality are weighted and summed according to the weight coefficients to generate a fusion feature representation under the corresponding spatiotemporal segment.

7. The method according to claim 1, characterized in that, The process of outputting filling parameter commands based on the policy network to control the filling process in a closed loop according to the continuous state field and physical constraints includes: The continuous state field is input into the policy network as the environment state. Based on the physical constraints, a constraint function is constructed to limit the strategy search space; A set of candidate filling parameters is generated by forward reasoning about the environment state through a policy network. The set of candidate filling parameters is filtered by applying constraint functions to obtain feasible solutions that satisfy the constraints; The feasible code is decoded into filling parameter instructions; and the filling parameter instructions are fed back to the filling equipment.

8. The method according to claim 1, characterized in that, The process of continuously monitoring and controlling the composite top co-filling surface by performing online incremental learning based on closed-loop feedback includes: The actual working condition data after the filling equipment is executed is compared with the continuous state field to calculate the monitoring error and deviation. Adjusting the weight parameters of the strategy network and neural network model based on errors and biases enables online incremental training; The updated model is reused for feature extraction, state field prediction, and filling parameter instruction generation; The incremental learning and closed-loop control process is executed cyclically, enabling the system to adaptively optimize the filling process during continuous monitoring, thereby achieving dynamic control and stable monitoring of the composite top co-filling surface.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computing program, it implements the method of any one of claims 1-8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-8.

Images