A Multi-Task Early Warning Method and System for Deep Hard Rockburst Based on Spatiotemporal Network of Dynamic and Static Stress Edge Attribute Enhancement Graph

By constructing a spatiotemporal network of dynamic and static stress edge attribute enhancement graphs, and combining it with edge enhancement graph attention networks and temporal convolutional branches, the shortcomings of existing rockburst early warning methods in integrating microseismic time series, excavation disturbance, and rock mechanics heterogeneity are addressed, achieving efficient prediction and accurate early warning of rockburst risks.

CN123116785APending Publication Date: 2026-11-17SHAOXING UNIVERSITY +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611218965.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-12
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing rockburst early warning methods cannot simultaneously integrate microseismic time-series evolution, excavation dynamic disturbance, rock mass mechanics heterogeneity, and principal stress direction constraints, resulting in inaccurate rockburst early warning.

Method used

A spatiotemporal network based on dynamic and static stress edge attribute enhancement graph is constructed. Spatial coupling modeling is performed through edge enhancement graph attention network. Local temporal mutations are captured by combining temporal convolutional branches. Adaptive fusion is performed through gating fusion mechanism to output the multi-task early warning results of rockburst within the future prediction window.

Benefits of technology

It improves the stability and accuracy of rockburst risk prediction, can adapt to the prediction of complex deformation stages, enhances the ability to express the mechanical heterogeneity of rock slopes, and provides a basis for engineering geologists to understand the information propagation relationship of monitoring points.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116785A_ABST
    Figure CN123116785A_ABST
Patent Text Reader

Abstract

This invention discloses a multi-task early warning method and system for deep hard rockburst based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graphs. The method includes: acquiring multi-source data of deep underground engineering; constructing a rockburst early warning graph structure based on the multi-source data; performing spatial coupling modeling through an edge-enhanced graph attention network based on the rockburst early warning graph structure to obtain a spatial coupling representation; capturing local temporal mutations through temporal convolution branches based on temporal sequence information in the multi-source data to obtain a local temporal mutation representation; obtaining a fused temporal representation based on the spatial coupling representation and the local temporal mutation representation; and outputting the multi-task early warning result of rockburst within the future prediction window based on the fused temporal representation through global temporal coding and multi-task prediction. This invention can solve the problem that existing rockburst early warning methods are unable to simultaneously integrate microseismic temporal evolution, excavation dynamic disturbance, rock mass mechanics heterogeneity, and principal stress direction constraints.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the fields of deep underground engineering disaster monitoring and early warning, microseismic monitoring data processing, rock mass mechanical parameter field modeling and artificial intelligence prediction technology, and in particular relates to a multi-task early warning method and system for deep hard rock bursts based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graph. Background Technology

[0002] Deep hard rock tunnels, mine roadways, underground powerhouses, and deeply buried transportation tunnels are prone to rockburst hazards under the influence of high ground stress, excavation unloading, blasting disturbance, and structural surface control. Rockbursts are characterized by their suddenness, rapid energy release, localized concentrated damage, and high difficulty in early warning, seriously threatening the safety of construction personnel, equipment, and underground engineering structures. Early identification and graded warning of rockbursts are crucial issues for safe construction and intelligent operation and maintenance of deep underground engineering projects.

[0003] Existing rockburst early warning methods mainly include empirical criteria, microseismic threshold methods, numerical simulation methods, and machine learning prediction methods. Empirical criteria typically judge rockburst tendency based on indicators such as rock strength, maximum principal stress, elastic energy index, and brittleness index, which can reflect certain mechanical mechanisms, but it is difficult to integrate construction disturbance and microseismic time-series evolution information in real time. Microseismic threshold methods judge risk status by monitoring changes in indicators such as the number of microseismic events, released energy, apparent volume, energy index, and b-value, but the threshold usually depends on human experience and is difficult to adapt to different surrounding rock conditions and excavation stages.

[0004] Deep learning methods can learn nonlinear patterns from large amounts of monitoring data, but traditional recurrent neural networks or convolutional neural networks often use single points or regular sequences as input, making it difficult to express the irregular spatial relationships between the microseismic sensor network around the underground cavern, the surrounding rock spatial grid, and the tunnel face advancement. Although existing graph neural network methods can express the topological relationships between sensors or spatial grids, they often rely mainly on Euclidean distance, adjacency relationships, or historical signal similarity when constructing the graph, without fully incorporating engineering mechanics factors such as high ground stress, heterogeneous rock stiffness, and the propagation path of excavation disturbances.

[0005] The rockburst incubation process is essentially closely related to the accumulation of high ground stress energy, excavation unloading disturbance, heterogeneity of the surrounding rock, and rapid penetration of local cracks. For spatially close units, if weak interlayers or stress release channels exist between them, the energy transfer between them may not be strong; for spatially distant units, if they are located within the same high principal stress channel or hard, brittle, and intact surrounding rock zone, they may have a stronger disturbance response correlation. Therefore, relying solely on geometric distance to construct diagrams is insufficient to accurately reflect the true propagation mechanism of microseismic activity and rockburst risk in deep hard rock.

[0006] Furthermore, pre-rockburst microseismic activity often exhibits phased characteristics, such as a gradual increase in the number of events, accelerated accumulation of released energy, a decrease in the b-value, a sudden increase after a local "quiet period," spatial migration of events, or concentration of high-energy events near the excavation face. Existing methods typically struggle to simultaneously characterize these short-term abrupt changes, long-term trends, spatial migration, and mechanical constraints. Therefore, a multi-task rockburst early warning technology solution is urgently needed that can integrate microseismic sequences, excavation disturbances, SMRM rock mass parameter fields, principal stress fields, and irregular topological relationships between sensors and spatial grids. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a multi-task early warning method and system for deep hard rock bursts based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graphs. This invention can solve the problem that existing rock burst early warning methods cannot simultaneously integrate microseismic temporal evolution, excavation dynamic disturbance, rock mass mechanics heterogeneity, and principal stress direction constraints.

[0008] To achieve the above objectives, this invention provides a multi-task early warning method for deep hard rock bursts based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graphs, comprising: Acquire multi-source data for deep underground engineering; Based on the multi-source data, a rockburst early warning map structure is constructed, wherein the rockburst early warning map structure includes graph nodes and graph edges; Based on the rockburst early warning map structure, spatial coupling modeling is performed using an edge-enhanced graph attention network to obtain a spatial coupling representation; Based on the temporal sequence information in the multi-source data, local temporal mutations are captured through temporal convolutional branches to obtain local temporal mutation representations. Based on the spatial coupling representation and the local temporal abrupt change representation, an adaptive fusion is performed through a gated fusion mechanism to obtain the fused temporal representation. Based on the fused temporal representation, the rockburst multi-task early warning result within the future prediction window is output through global temporal coding and multi-task prediction.

[0009] Optionally, the multi-source data for the deep underground engineering includes: microseismic sequence data, excavation disturbance data, blasting or mechanical operation disturbance data, SMRM rock mass parameter field, principal stress field, spatial geometric data, and temporal characteristic data.

[0010] Optionally, constructing the rockburst early warning map structure based on the multi-source data includes: Microseismic sensors, microseismic event cluster centers, or surrounding rock spatial grids are used as graph nodes, and graph edges are established based on the spatial proximity between nodes, microseismic propagation paths, tunnel face advance directions, or directed fully connected relationships without self-loops. Based on the graph nodes and graph edges, construct the rockburst early warning graph structure.

[0011] Optionally, based on the rockburst early warning map structure, spatial coupling modeling is performed using an edge-enhanced graph attention network to obtain the spatial coupling representation, including: Based on the mechanical parameters and disturbance parameters corresponding to the nodes at both ends of the graph, construct the dynamic and static stress edge properties; Based on the multi-source data, multimodal feature encoding is performed on each graph node to obtain the initial node representation; Based on the initial node representation, rockburst early warning map structure, and dynamic and static stress edge properties, spatial coupling modeling is performed using an edge-enhanced graph attention network to obtain the spatial coupling representation.

[0012] Optionally, based on the mechanical parameters and disturbance parameters corresponding to the nodes at both ends of the graph edge, the dynamic and static stress edge properties are constructed, including: Based on the equivalent elastic modulus of the rock mass, the magnitude and direction of the principal stress, the distance between the excavation faces and the disturbance intensity at the corresponding spatial locations of the nodes at both ends of the graph edge, calculate the static mechanical edge properties and the dynamic disturbance edge properties. Based on the static mechanical edge attributes and dynamic disturbance edge attributes, the principal stress projection coefficient and the blasting or mechanical disturbance attenuation coefficient are used as adjustment terms of the dynamic disturbance edge attributes to obtain the dynamic and static stress edge attributes; wherein, the principal stress projection coefficient is used to characterize the consistency between the edge direction and the direction of the maximum principal stress, and the blasting or mechanical disturbance attenuation coefficient is used to characterize the intensity of the effect of construction disturbance after attenuation with spatial distance and time interval.

[0013] Optionally, based on the temporal sequence information in the multi-source data, local temporal mutation capture is performed through temporal convolutional branches to obtain local temporal mutation representations, including: The microseismic sequence and disturbance sequence of each graph node within the historical time window are input into the temporal convolution branch, which includes at least two one-dimensional convolutional layers, a nonlinear activation layer and a regularization layer. The microseismic calm period, event surge period, energy surge and disturbance response features are extracted to obtain the local temporal abrupt change representation.

[0014] Optionally, based on the spatial coupling representation and the local temporal abrupt change representation, adaptive fusion is performed through a gating fusion mechanism to obtain the fused temporal representation, including: The spatially coupled representation and the local temporal abrupt change representation are concatenated and then input into a gating network to obtain the gating weights. The spatial coupling representation and the local temporal abrupt change representation are weighted and fused according to the gating weights to obtain the fused temporal representation.

[0015] Optionally, based on the fused temporal representation, the output of the rockburst multi-task early warning result within the future prediction window, through global temporal coding and multi-task prediction, includes: The fused temporal representation is input into a bidirectional long short-term memory network, and the global temporal context is extracted along the forward and backward directions of the historical window to obtain the hidden state at each historical moment. The temporal attention weights are calculated based on the correlation between each hidden state and the rockburst early warning task, and the hidden states are weighted and aggregated to obtain the context representation of key moments. The key moment context representation is input into the multi-task prediction head, and the shared representation is mapped to at least two early warning results among the frequency of microseismic events, released energy, risk spatial location, and rockburst risk level within the future prediction window.

[0016] The present invention also provides a multi-task early warning system for rockburst in deep hard rock based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graph, comprising: a data acquisition module, a graph structure construction module, a spatiotemporal coupling module, and a multi-task prediction output module; The data acquisition module is used to acquire multi-source data from deep underground engineering projects; The graph structure construction module is used to construct a rockburst early warning graph structure based on the multi-source data, wherein the rockburst early warning graph structure includes graph nodes and graph edges; The spatiotemporal coupling module is used to perform spatial coupling modeling based on the rockburst early warning map structure through an edge-enhanced graph attention network to obtain a spatial coupling representation; according to the temporal sequence information in the multi-source data, it performs local temporal mutation capture through temporal convolutional branches to obtain a local temporal mutation representation; and according to the spatial coupling representation and the local temporal mutation representation, it performs adaptive fusion through a gating fusion mechanism to obtain a fused temporal representation. The multi-task prediction output module is used to output the rockburst multi-task early warning result within the future prediction window based on the fused temporal representation, through global temporal coding and multi-task prediction.

[0017] Compared with the prior art, the present invention has the following advantages and technical effects: 1. This invention uses monitoring points as graph nodes, which is suitable for scenarios where monitoring points on rock slopes are sparse and irregularly arranged, thus avoiding the dependence of conventional convolutional models on regular grid data.

[0018] 2. This invention constructs the normalized difference of elastic modulus as an edge attribute and introduces a graph attention propagation process, so that when the model aggregates monitoring point information, it considers not only geometric relationships but also differences in rock mass stiffness, thereby improving the ability to express the heterogeneity of rock slope mechanics.

[0019] 3. This invention, through graph space branching, temporal convolutional branching, and gating fusion mechanisms, simultaneously characterizes the spatial coupling relationship between monitoring points, local short-term disturbances, and long-term temporal evolution, thereby improving the prediction stability of complex deformation stages.

[0020] 4. This invention enhances the response capability to the most recent deformation state while preserving the global spatiotemporal context through a bidirectional long short-term memory network, a temporal attention module, and a short-circuit correction branch, which is beneficial for future multi-step displacement prediction.

[0021] 5. The input, edge attributes, and attention propagation mechanism of this invention all have clear engineering implications and can provide a basis for engineering geologists to understand the information propagation relationship between monitoring points. Attached Figure Description

[0022] 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: Figure 1 This is a flowchart of a multi-task early warning method for deep hard rock bursts based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graph according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the relationship between the microseismic sensor network, the surrounding rock spatial grid, and the tunnel face advancement in an embodiment of the present invention. Figure 3 This is a schematic diagram of the attention propagation result of the elastic modulus edge property enhancement graph according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the edge-enhanced graph attention network, temporal convolutional branch, and gated fusion module structure according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the rockburst multi-task early warning output according to an embodiment of the present invention. Detailed Implementation

[0023] 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.

[0024] 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.

[0025] This embodiment proposes a multi-task early warning method for deep hard rock bursts based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graphs, such as... Figure 1 As shown, the specific steps include: Acquire multi-source data for deep underground engineering; Based on the multi-source data, a rockburst early warning map structure is constructed, wherein the rockburst early warning map structure includes graph nodes and graph edges; Based on the rockburst early warning map structure, spatial coupling modeling is performed using an edge-enhanced graph attention network to obtain a spatial coupling representation; Based on the temporal sequence information in the multi-source data, local temporal mutations are captured through temporal convolutional branches to obtain local temporal mutation representations. Based on the spatial coupling representation and the local temporal abrupt change representation, an adaptive fusion is performed through a gated fusion mechanism to obtain the fused temporal representation. Based on the fused temporal representation, the rockburst multi-task early warning result within the future prediction window is output through global temporal coding and multi-task prediction.

[0026] Furthermore, the multi-source data of the deep underground engineering includes: microseismic sequence data, excavation disturbance data, blasting or mechanical operation disturbance data, SMRM rock mass parameter field, principal stress field, spatial geometric data, and temporal characteristic data.

[0027] Specifically, S1: Acquire multi-source monitoring and mechanical data for deep underground engineering. The multi-source monitoring and mechanical data includes microseismic sequence data, excavation disturbance data, blasting or mechanical operation disturbance data, SMRM rock mass parameter field, principal stress field, spatial geometric data, and temporal characteristic data.

[0028] Microseismic sequence data includes one or more of the following: microseismic event time, event spatial coordinates, number of events, released energy, magnitude, peak amplitude, dominant frequency, apparent volume, energy index, cumulative energy, b-value, and sensor channel characteristics; excavation disturbance data includes one or more of the following: face location, daily excavation advance, excavation direction, distance from nodes, excavation method, and support sequence; blasting or mechanical operation disturbance data includes one or more of the following: blast charge, blasting time, blasting vibration velocity, equipment impact frequency, impact energy, cutterhead thrust, torque, and operation duration; SMRM rock mass parameter field includes one or more of the following: rock mass equivalent elastic modulus, equivalent Poisson's ratio, equivalent cohesion, equivalent internal friction angle, brittleness index, failure probability, and risk index; principal stress field includes one or more of the following: maximum principal stress, minimum principal stress, principal stress direction, stress concentration factor, and confining pressure state.

[0029] Furthermore, based on the aforementioned multi-source data, the structure for constructing a rockburst early warning map includes: Microseismic sensors, microseismic event cluster centers, or surrounding rock spatial grids are used as graph nodes, and graph edges are established based on the spatial proximity between nodes, microseismic propagation paths, tunnel face advance directions, or directed fully connected relationships without self-loops. Based on the graph nodes and graph edges, construct the rockburst early warning graph structure.

[0030] Specifically, S2: Constructing the structure of a rockburst early warning map, as follows: Figure 2As shown, microseismic sensors, microseismic event cluster centers, or three-dimensional spatial grids of the surrounding rock around the underground cavern are used as graph nodes; graph edges are established based on the spatial proximity between nodes, microseismic propagation paths, tunnel face advance directions, or directed fully connected relationships without self-loops; edge attributes are constructed for each graph edge, including static mechanical edge attributes and dynamic disturbance edge attributes.

[0031] Furthermore, based on the aforementioned rockburst early warning map structure, spatial coupling modeling is performed using an edge-enhanced graph attention network to obtain the spatial coupling representation, including: Based on the mechanical parameters and disturbance parameters corresponding to the nodes at both ends of the graph, construct the dynamic and static stress edge properties; Based on the multi-source data, multimodal feature encoding is performed on each graph node to obtain the initial node representation; Based on the initial node representation, rockburst early warning map structure, and dynamic and static stress edge properties, spatial coupling modeling is performed using an edge-enhanced graph attention network to obtain the spatial coupling representation.

[0032] Specifically, S3: Constructing dynamic and static stress edge properties such as Figure 3 As shown. For an edge pointing from node i to node j, an edge attribute vector is constructed based on the equivalent elastic modulus of the rock mass, the magnitude of the principal stresses, the direction of the principal stresses, the distance between the excavation faces, and the disturbance intensity at the corresponding spatial locations of nodes i and j. Preferably, the edge attribute vector includes the distance between nodes, the difference in elastic modulus, the difference in maximum principal stress, the difference in minimum principal stress, the angle between the principal stress directions, the projection of the principal stresses along the edge direction, the difference in excavation disturbance, and the blasting vibration attenuation coefficient.

[0033] S4: Constructing rockburst multi-task tags, such as... Figure 4 As shown, a multi-task prediction objective is constructed based on microseismic events and on-site rockburst records within a future preset prediction window. The prediction objective includes one or more of the following: the frequency of microseismic events in the future window, the total or maximum released energy in the future window, the future high-risk spatial grid or high-risk sensor area, and the rockburst risk level.

[0034] S5: Perform node multimodal input encoding as follows Figure 4 As shown, the microseismic statistical features, waveform features, excavation disturbance features, blasting or mechanical operation disturbance features, SMRM parameters, and time features of each node within the historical time window are linearly mapped or embedded and encoded, and then fused into the initial node representation in a unified latent space.

[0035] S6: Perform attention modeling of dynamic and static stress edge enhancement graphs, such as... Figure 4As shown, the initial node representation, graph structure, and dynamic and static stress edge attributes are input into the edge-enhanced graph attention network. When calculating the information propagation weights between nodes, the graph attention network simultaneously incorporates source node features, target node features, and edge attributes, enabling the model to dynamically adjust spatial information aggregation relationships based on rock mechanics heterogeneity, principal stress directions, and construction disturbance intensity.

[0036] Furthermore, based on the mechanical parameters and disturbance parameters corresponding to the nodes at both ends of the graph edge, the dynamic and static stress edge properties are constructed as follows: Based on the equivalent elastic modulus of the rock mass, the magnitude and direction of the principal stress, the distance between the excavation faces and the disturbance intensity at the corresponding spatial locations of the nodes at both ends of the graph edge, calculate the static mechanical edge properties and the dynamic disturbance edge properties. Based on the static mechanical edge attributes and dynamic disturbance edge attributes, the principal stress projection coefficient and the blasting or mechanical disturbance attenuation coefficient are used as adjustment terms for the dynamic disturbance edge attributes to obtain the dynamic and static stress edge attributes. Among them, the principal stress projection coefficient is used to characterize the consistency between the edge direction and the direction of the maximum principal stress, and the blasting or mechanical disturbance attenuation coefficient is used to characterize the intensity of the effect of construction disturbance after it has attenuated with spatial distance and time interval, so that the edge attributes can simultaneously reflect the heterogeneity of rock mechanics, the constraint of principal stress direction, and the influence of excavation disturbance propagation.

[0037] Furthermore, based on the temporal sequence information in the multi-source data, local temporal mutation capture is performed through temporal convolutional branches to obtain local temporal mutation representations, including: The microseismic sequence and disturbance sequence of each graph node within the historical time window are input into the temporal convolution branch, which includes at least two one-dimensional convolutional layers, a nonlinear activation layer and a regularization layer. The microseismic calm period, event surge period, energy surge and disturbance response features are extracted to obtain the local temporal abrupt change representation.

[0038] Specifically, S7: Captures timing abrupt changes in rockburst precursors, such as... Figure 4 As shown, the historical microseismic sequences and disturbance sequences of each node are input into a temporal convolutional branch to extract local temporal patterns such as short-term calm periods, event surge periods, energy spikes, and disturbance responses. The temporal convolutional branch includes at least two one-dimensional convolutional layers, a nonlinear activation layer, and a regularization layer.

[0039] Furthermore, based on the spatial coupling representation and the local temporal abrupt change representation, adaptive fusion is performed through a gated fusion mechanism to obtain the fused temporal representation, including: The spatially coupled representation and the local temporal abrupt change representation are concatenated and then input into a gating network to obtain the gating weights. The spatial coupling representation and the local temporal abrupt change representation are weighted and fused according to the gating weights to obtain the fused temporal representation.

[0040] Specifically, S8: performs spatial-temporal dual-branch gating fusion, such as... Figure 4 As shown, the spatially coupled representation output by the edge-enhanced graph attention network and the local mutation representation output by the temporal convolution branch are concatenated and input into the gating network to generate gating weights. The two types of representations are then adaptively fused based on the gating weights.

[0041] Furthermore, based on the fused temporal representation, through global temporal coding and multi-task prediction, the output of the rockburst multi-task early warning results within the future prediction window includes: The fused temporal representation is input into a bidirectional long short-term memory network, and the global temporal context is extracted along the forward and backward directions of the historical window to obtain the hidden state at each historical moment. The temporal attention weights are calculated based on the correlation between each hidden state and the rockburst early warning task, and the hidden states are weighted and aggregated to obtain the context representation of key moments. The key moment context representation is input into the multi-task prediction head, and the shared representation is mapped to at least two early warning results among the frequency of microseismic events, released energy, risk spatial location, and rockburst risk level within the future prediction window.

[0042] Specifically, S9: performs global timing encoding and multi-task early warning output, such as... Figure 5 As shown, the fused temporal representation is first input into a bidirectional long short-term memory network, and the global temporal context is extracted along the forward and backward directions of the historical window to obtain the hidden state at each historical moment. Then, the temporal attention module calculates the temporal attention weights based on the correlation between each hidden state and the rockburst early warning task, and the hidden states are weighted and aggregated to obtain the key moment context representation. Finally, the key moment context representation is input into the multi-task prediction head, and the shared representation is mapped to the frequency of microseismic events, released energy, risk spatial location, and rockburst risk level, respectively.

[0043] S10: Perform model training and generate early warning results. Divide the model into training, validation, and test sets according to time sequence, and fit a normalizer only on the training set; use hard sample resampling, class imbalance weighting, or focus loss to strengthen the training of high-risk samples; denormalize and map the model output to generate rockburst early warning levels, risk locations, and recommended handling results within the future prediction window.

[0044] This embodiment provides a multi-task early warning method for rockbursts in deep hard rock based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graphs. The research object is a hard rock cavern in a deep buried tunnel, deep mine roadway, or underground powerhouse. Multiple microseismic sensors are deployed around the cavern, and the surrounding rock area is divided into several three-dimensional spatial grids according to the engineering spatial coordinates.

[0045] Step 1: Acquire multi-source monitoring and mechanical data: In one embodiment, the microseismic monitoring system outputs an event catalog and channel characteristics in real time. The event catalog includes event occurrence time, event location coordinates, released energy, magnitude, peak amplitude, dominant frequency, apparent volume, and energy index. The continuous time is divided into time slices of fixed length, such as 10 minutes, 30 minutes, or 1 hour. Within each time slice, the number of events, cumulative released energy, maximum released energy, average dominant frequency, b-value, energy growth rate, and event spatial migration distance are statistically analyzed for each spatial grid or sensor neighborhood.

[0046] Simultaneously, disturbance data such as the face location, excavation advance, excavation direction, blasting time, explosive charge, blasting vibration velocity, mechanical impact frequency, and equipment operating intensity are obtained from the construction management system. The equivalent elastic modulus, equivalent Poisson's ratio, strength parameters, failure probability, and risk indicators corresponding to the surrounding rock spatial grid are obtained from the SMRM model or numerical simulation results; the maximum principal stress, minimum principal stress, and principal stress direction are obtained from the in-situ stress test or numerical simulation results.

[0047] Step 2: Construct graph nodes and graph edges: If microseismic sensors are used as nodes, each node corresponds to a sensor and its neighborhood space; if surrounding rock spatial grids are used as nodes, each node corresponds to a three-dimensional grid element. For nodes i and j, a directed edge is established if their spatial distance is less than a preset threshold, or if they are within the same principal stress dominance propagation direction, or if they are both within the tunnel face disturbance influence zone. Alternatively, a directed complete graph without self-loops can be constructed, allowing each node to automatically learn the association strength with other nodes based on an attention mechanism.

[0048] Step 3: Construct the dynamic and static stress edge properties: For each edge, construct edge attributes containing both static mechanical information and dynamic construction disturbance information. Static mechanical information includes spatial distance, difference in SMRM equivalent elastic modulus, difference in maximum principal stress, difference in minimum principal stress, angle between principal stress directions, and projection of principal stresses along the edge direction. Dynamic construction disturbance information includes the distance difference from the nodes at both ends of the edge to the tunnel face, the angle between the excavation face advance direction and the edge direction, the blasting vibration velocity attenuation coefficient, and the difference in mechanical impact disturbance intensity. The principal stress projection coefficient can be expressed as: ; in, The average value of the maximum principal stress at node i and node j. Let be the average of the minimum principal stresses at node i and node j. It is the angle between the direction of the node connection and the direction of the maximum principal stress.

[0049] The attenuation coefficient of blasting or mechanical disturbance can be expressed as: ; in, Let t be the intensity of the disturbance caused by blasting or mechanical operation. Let be the distance from the disturbance source to the midpoint of the edge. This is the attenuation coefficient.

[0050] Step 4: Generate model input samples: Let the historical time window length be L, and the prediction window length be H. For each node, a node sequence feature matrix is ​​formed within L historical time slices. The features include the number of microseismic events, cumulative energy, maximum energy, b-value, apparent volume, dominant frequency, energy growth rate, quiescence index, excavation distance, blasting disturbance intensity, SMRM equivalent elastic modulus, principal stress, and failure probability. The quiescence index is used to characterize the short-term reduction in microseismic activity that may occur before rockburst, and can be constructed by the ratio of the current sliding window event rate to the event rate of a longer historical window.

[0051] The silence index can be expressed as: ; in, Let i be the average number of events within the short window. Let i be the average number of events within a long window. It is a stable term. If A relatively large value indicates that local microseismic activity is lower than historical levels, and may be in a phase of energy accumulation or short-term calm.

[0052] Step 5: Attention space modeling of edge-enhanced graph: After mapping the node sequence features to a unified latent space, they are input into the edge-enhanced graph attention layer. The edge-enhanced graph attention layer uses both node features and dynamic and static stress edge attributes to calculate the attention weights. If two nodes are highly correlated in the principal stress transmission direction, have similar rock mass stiffness, and are affected by the same excavation disturbance, the model can assign them higher propagation weights; if two nodes are geometrically close but have significantly different mechanical properties or the disturbance propagation is hindered, the model can reduce their propagation weights.

[0053] Step 6: Temporal mutation modeling and gating fusion: In addition to the spatial graph branch, the model incorporates a temporal convolution branch in parallel. This branch performs convolution operations on the sequence of microseismic events, released energy, b-value, quiescent index, and disturbance intensity to capture local short-term abrupt change patterns. The gating fusion module generates gating weights based on the graph branch output and the temporal convolution branch output, thereby automatically adjusting the model's dependence on spatial stress correlations and local microseismic abrupt changes at different stages.

[0054] Step 7, Multi-task prediction: The fused temporal representation is fed into a bidirectional long short-term memory network to capture long-term evolutionary relationships within the historical window. A temporal attention module assigns higher weights to key time slices preceding rockbursts, forming a global contextual representation. The multi-task prediction head outputs the frequency of microseismic events, released energy, risk locations, and rockburst risk levels for the next H time slices.

[0055] In one embodiment, the rockburst risk level can be divided into four levels: normal, concern, warning, and danger. If the predicted event frequency and released energy both increase rapidly, and the high-risk spatial location is close to the working face or the area of ​​high principal stress concentration, a higher risk level is output; if the prediction results show that the number of events is low and there is no significant energy accumulation, a normal or concern level is output.

[0056] Step 8: Training Optimization The training phase divides the dataset into training, validation, and test sets according to time sequence. Since high-risk rockburst samples are typically few in number, class weighting, focus loss, or resampling of difficult samples can be applied to these samples. For event frequency tasks, mean squared error loss, Poisson loss, or negative binomial loss can be used; for energy tasks, mean squared error loss can be applied after taking the logarithm of the energy value; for location and grade tasks, cross-entropy loss or focus loss can be used. Early stopping, gradient pruning, and learning rate scheduling are used during training to improve stability.

[0057] Step 9: Output the early warning results: During model deployment, data from the microseismic monitoring system, construction system, blasting monitoring system, and SMRM parameter field are read in real time, and map node features and dynamic edge attributes are updated according to fixed time slices. The model outputs the event frequency, released energy, risk location, and rockburst level for the future prediction window, and marks high-risk areas on the 3D cavern model or sensor network map. The system can output response suggestions based on the risk level, such as strengthening monitoring, adjusting excavation advance, optimizing blasting parameters, implementing pre-decompression, or temporary evacuation.

[0058] This embodiment also provides a multi-task early warning system for rockburst in deep hard rock based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graph, including: a data acquisition module, a graph structure construction module, a spatiotemporal coupling module, and a multi-task prediction output module; The data acquisition module is used to acquire multi-source data from deep underground engineering projects; The graph structure construction module is used to construct a rockburst early warning graph structure based on the multi-source data, wherein the rockburst early warning graph structure includes graph nodes and graph edges; The spatiotemporal coupling module is used to perform spatial coupling modeling based on the rockburst early warning map structure through an edge-enhanced graph attention network to obtain a spatial coupling representation; according to the temporal sequence information in the multi-source data, it performs local temporal mutation capture through temporal convolutional branches to obtain a local temporal mutation representation; and according to the spatial coupling representation and the local temporal mutation representation, it performs adaptive fusion through a gating fusion mechanism to obtain a fused temporal representation. The multi-task prediction output module is used to output the rockburst multi-task early warning result within the future prediction window based on the fused temporal representation, through global temporal coding and multi-task prediction.

[0059] Specifically, this embodiment provides a multi-task early warning system for rockburst in deep hard rock based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graph. It includes: a data acquisition module, a graph construction module, a dynamic and static stress edge attribute construction module, a node feature encoding module, an edge enhancement spatial modeling module, a temporal mutation capture module, a gating fusion module, a multi-task prediction module, a training and optimization module, and an early warning output module.

[0060] The data acquisition module is used to acquire microseismic sequence data, excavation disturbance data, blasting or mechanical operation disturbance data, SMRM rock mass parameter field, principal stress field, spatial geometric data, and temporal characteristic data.

[0061] The graph construction module is used to construct graph nodes from microseismic sensors, microseismic event cluster centers, or surrounding rock 3D spatial meshes, and to establish graph edges based on spatial proximity, principal stress propagation relationships, or fully connected relationships.

[0062] The dynamic and static stress edge attribute construction module is used to calculate the distance between nodes, the difference in SMRM equivalent elastic modulus, the difference in principal stress magnitude, the angle between principal stress directions, the principal stress projection coefficient, excavation disturbance attributes, and blasting or mechanical operation disturbance attributes.

[0063] The node feature encoding module is used to encode and fuse microseismic statistical features, waveform features, excavation disturbance features, SMRM parameters, principal stress parameters, and time features.

[0064] The edge-enhanced spatial modeling module is used to learn the spatial risk propagation relationships between different nodes through an edge-enhanced graph attention network.

[0065] The temporal mutation capture module is used to capture microseismic calm periods, surge periods, energy surges, and disturbance response characteristics through temporal convolutional branches.

[0066] The gating fusion module is used to adaptively fuse spatial risk propagation representations and local temporal mutation representations.

[0067] The multi-task prediction module is used to output the frequency of microseismic events, released energy, risk spatial location, and rockburst risk level within the future prediction window.

[0068] The training optimization module is used to perform time-sequential partitioning, data standardization, imbalanced sample handling, hard sample resampling, and model parameter optimization.

[0069] The early warning output module is used to de-standardize the prediction results, map the risk level, and visualize the results, and output on-site handling suggestions.

[0070] 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 deep hard rock rockburst multi-task early warning method based on dynamic and static stress edge attribute enhancement of spatiotemporal network, characterized in that, include: Acquire multi-source data for deep underground engineering; Based on the multi-source data, a rockburst early warning map structure is constructed, wherein the rockburst early warning map structure includes graph nodes and graph edges; Based on the rockburst early warning map structure, spatial coupling modeling is performed using an edge-enhanced graph attention network to obtain a spatial coupling representation; Based on the temporal sequence information in the multi-source data, local temporal mutations are captured through temporal convolutional branches to obtain local temporal mutation representations. Based on the spatial coupling representation and the local temporal abrupt change representation, an adaptive fusion is performed through a gated fusion mechanism to obtain the fused temporal representation. Based on the fused temporal representation, the rockburst multi-task early warning result within the future prediction window is output through global temporal coding and multi-task prediction.

2. The deep hard rock rockburst multi-task early warning method based on dynamic and static stress edge attribute enhanced graph space-time network according to claim 1, characterized in that, The multi-source data for the deep underground engineering includes: microseismic sequence data, excavation disturbance data, blasting or mechanical operation disturbance data, SMRM rock mass parameter field, principal stress field, spatial geometric data, and temporal characteristic data.

3. The multi-task early warning method for deep hard rock bursts based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graphs as described in claim 1, characterized in that, Based on the aforementioned multi-source data, the structure for constructing a rockburst early warning map includes: Microseismic sensors, microseismic event cluster centers, or surrounding rock spatial grids are used as graph nodes, and graph edges are established based on the spatial proximity between nodes, microseismic propagation paths, tunnel face advance directions, or directed fully connected relationships without self-loops. Based on the graph nodes and graph edges, construct the rockburst early warning graph structure.

4. The multi-task early warning method for deep hard rock bursts based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graphs as described in claim 1, characterized in that, Based on the rockburst early warning map structure, spatial coupling modeling is performed using an edge-enhanced graph attention network to obtain the spatial coupling representation, including: Based on the mechanical parameters and disturbance parameters corresponding to the nodes at both ends of the graph, construct the dynamic and static stress edge properties; Based on the multi-source data, multimodal feature encoding is performed on each graph node to obtain the initial node representation; Based on the initial node representation, rockburst early warning map structure, and dynamic and static stress edge properties, spatial coupling modeling is performed using an edge-enhanced graph attention network to obtain the spatial coupling representation.

5. A multi-task early warning method for deep hard rock bursts based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graphs, as described in claim 4, is characterized in that... Based on the mechanical parameters and disturbance parameters corresponding to the nodes at both ends of the graph edge, the dynamic and static stress edge properties are constructed as follows: Based on the equivalent elastic modulus of the rock mass, the magnitude and direction of the principal stress, the distance between the excavation faces and the disturbance intensity at the corresponding spatial locations of the nodes at both ends of the graph edge, calculate the static mechanical edge properties and the dynamic disturbance edge properties. Based on the static mechanical edge attributes and dynamic disturbance edge attributes, the principal stress projection coefficient and the blasting or mechanical disturbance attenuation coefficient are used as adjustment terms of the dynamic disturbance edge attributes to obtain the dynamic and static stress edge attributes; wherein, the principal stress projection coefficient is used to characterize the consistency between the edge direction and the direction of the maximum principal stress, and the blasting or mechanical disturbance attenuation coefficient is used to characterize the intensity of the effect of construction disturbance after attenuation with spatial distance and time interval.

6. A multi-task early warning method for deep hard rock bursts based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graphs, as described in claim 1, is characterized in that... Based on the temporal sequence information in the multi-source data, local temporal mutations are captured through temporal convolutional branches, and the representation of local temporal mutations is obtained, including: The microseismic sequence and disturbance sequence of each graph node within the historical time window are input into the temporal convolution branch, which includes at least two one-dimensional convolutional layers, a nonlinear activation layer and a regularization layer. The microseismic calm period, event surge period, energy surge and disturbance response features are extracted to obtain the local temporal abrupt change representation.

7. A multi-task early warning method for deep hard rock bursts based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graphs, as described in claim 1, is characterized in that... Based on the spatial coupling representation and the local temporal abrupt change representation, adaptive fusion is performed through a gated fusion mechanism to obtain the fused temporal representation, including: The spatially coupled representation and the local temporal abrupt change representation are concatenated and then input into a gating network to obtain the gating weights. The spatial coupling representation and the local temporal abrupt change representation are weighted and fused according to the gating weights to obtain the fused temporal representation.

8. A multi-task early warning method for deep hard rock bursts based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graphs, as described in claim 1, is characterized in that... Based on the fused temporal representation, through global temporal coding and multi-task prediction, the output of rockburst multi-task early warning results within the future prediction window includes: The fused temporal representation is input into a bidirectional long short-term memory network, and the global temporal context is extracted along the forward and backward directions of the historical window to obtain the hidden state at each historical moment. The temporal attention weights are calculated based on the correlation between each hidden state and the rockburst early warning task, and the hidden states are weighted and aggregated to obtain the context representation of key moments. The key moment context representation is input into the multi-task prediction head, and the shared representation is mapped to at least two early warning results among the frequency of microseismic events, released energy, risk spatial location, and rockburst risk level within the future prediction window.

9. A multi-task early warning system for rockburst in deep hard rock based on a spatiotemporal network of dynamic and static stress edge attribute enhancement graph, used to implement the method described in any one of claims 1-8, characterized in that, include: The module consists of a data acquisition module, a graph structure construction module, a spatiotemporal coupling module, and a multi-task prediction output module. The data acquisition module is used to acquire multi-source data from deep underground engineering projects; The graph structure construction module is used to construct a rockburst early warning graph structure based on the multi-source data, wherein the rockburst early warning graph structure includes graph nodes and graph edges; The spatiotemporal coupling module is used to perform spatial coupling modeling based on the rockburst early warning map structure through an edge-enhanced graph attention network to obtain a spatial coupling representation; according to the temporal sequence information in the multi-source data, it performs local temporal mutation capture through temporal convolutional branches to obtain a local temporal mutation representation; and according to the spatial coupling representation and the local temporal mutation representation, it performs adaptive fusion through a gating fusion mechanism to obtain a fused temporal representation. The multi-task prediction output module is used to output the rockburst multi-task early warning result within the future prediction window based on the fused temporal representation, through global temporal coding and multi-task prediction.