Tunnel surrounding rock structure surface cross-cycle correlation analysis method and system based on graph memory

Abstract

The application relates to a tunnel surrounding rock structure surface cross-cycle correlation analysis method and system based on graph memory, which obtains a face image, a point cloud, geological records, water seepage, block falling and deformation monitoring data under continuous multiple excavation cycles, extracts joints, cracks, bedding, water seepage points, block falling areas and deformation anomaly units after unified expression, and constructs graph nodes, further constructs a single-cycle structure graph based on space adjacency, same-group joints, geometric continuation, local co-occurrence and risk propagation relationship; through cross-cycle node matching and graph memory network updating, continuous memory and dynamic correlation of the surrounding rock structure object are realized; based on the graph memory network, the same structure surface cross-cycle extension, crack transformation into a block falling area, water seepage point expansion into a water-rich zone and structure anomaly and monitoring anomaly coupling enhancement mode are identified, and risk accumulation evaluation and key attention area determination are carried out. The application can realize surrounding rock structure continuity identification, anomaly expansion path tracking and risk dynamic quantification.

Description

Technical Field

[0001] This invention relates to the field of tunnel and underground engineering technology, specifically to a method and system for cross-cycle correlation analysis of tunnel surrounding rock structural surfaces based on graph memory. Background Technology

[0002] With the continuous expansion of the construction scale of mountain tunnels, deep-buried tunnels and underground engineering projects under complex geological conditions, the identification of the surrounding rock structure at the working face and the early warning of construction risks have become important technical links to ensure excavation safety, optimize support parameters and improve the level of construction organization refinement. In common construction methods such as drill-and-blast method, mechanical excavation method and bench method, the surrounding rock structure usually appears in the form of joints, fissures, bedding, weak interlayers, seepage points, rockfall areas and local deformation anomalies. These structural anomalies do not exist independently in a single excavation cycle, but often show continuous appearance, spatial extension, mutual coupling and risk accumulation characteristics between adjacent or even multiple excavation cycles as the working face continues to advance.

[0003] In existing technologies, the identification of the surrounding rock condition at the tunnel face typically employs methods such as image recognition, point cloud reconstruction, geological sketching, radar detection, manual inspection, or monitoring data analysis to identify and interpret the distribution of joints and fissures, seepage locations, localized fractured areas, and deformation anomalies under the current cycle. These technologies play a certain role in extracting the surrounding rock condition for a single cycle, a single cross-section, or a single moment, and can provide a basic basis for on-site geological logging, construction judgment, and support adjustment.

[0004] However, most existing technologies focus on real-time identification or local analysis within the current excavation cycle. They typically store and process the structural identification results from different cycles separately, lacking a unified correlation modeling mechanism for multiple consecutive excavation cycles. As a result, it is difficult to effectively characterize the continuity, branching, and evolution trajectory of the same structural surface in successive cycles. For example, the same set of joints may continue to appear in multiple cycles, cracks may gradually expand and evolve into slab areas with excavation disturbance, and local seepage points may further connect and develop into water-rich zones. Existing methods often can only provide identification results for each cycle separately, making it difficult to determine whether they belong to the same geological structure or the same risk evolution chain.

[0005] Furthermore, existing technologies are insufficient in expressing the relationship between structural anomalies in surrounding rock and risk evolution. Objects such as joints, fissures, seepage, rockfalls, and deformation monitoring anomalies typically exhibit complex relationships including spatial adjacency, grouped distribution, continuous propagation, and risk transmission. However, existing methods primarily rely on isolated target detection, threshold judgment, or manual experience comparison, lacking a structured model that can simultaneously express "object-relationship-historical state." This makes it difficult to identify typical patterns during continuous excavation, such as the transformation of fissures into rockfall areas, the expansion of seepage points into water-rich zones, and the synergistic enhancement of structural anomalies and monitoring anomalies.

[0006] Furthermore, existing technologies typically lack a mechanism for continuous memory and dynamic updating of historical cycle information. As tunnel excavation progresses, the exposed surface of the tunnel face constantly changes. Structural features identified in the previous cycle may continue to appear, disappear locally, shift in direction, or expand in scale in the subsequent cycle. If there is no unified method for expressing cross-cycle memory, relying solely on manual review of images, recording, or discrete monitoring reports can easily lead to inaccurate judgments of structural continuity, difficulty in tracing abnormal expansion paths, and difficulty in quantifying risk development trends. This, in turn, affects the relevance and timeliness of subsequent decisions regarding the delineation of key areas of focus, the arrangement of advanced support, and the intensification of monitoring.

[0007] Therefore, how to address the issues of continuous manifestation, propagation, and risk accumulation of the tunnel face structure in different excavation cycles, and how to construct a method and system capable of unified coding, cross-cycle correlation, continuous memory, and dynamic analysis of joint, crack, bedding, seepage point, spalling area, and deformation monitoring results, thereby achieving continuous identification of the structural face, tracking of abnormal expansion paths, assessment of risk accumulation levels, and determination of subsequent key areas of concern, has become an urgent technical problem to be solved in this field. Summary of the Invention

[0008] Based on the above description, this invention provides a method and system for cross-cycle correlation analysis of tunnel surrounding rock structural surfaces based on graph memory, in order to solve the problems in the prior art such as fragmented identification results of surrounding rock structural surfaces, difficulty in characterizing the continuity relationship between different excavation cycles, difficulty in tracking abnormal expansion paths, difficulty in quantifying the risk accumulation process, and difficulty in accurately determining the key areas of concern in the future.

[0009] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0010] The method for cross-cycle correlation analysis of tunnel surrounding rock structural surfaces based on graph memory includes the following steps:

[0011] S1. Acquisition and unified representation of multi-source surrounding rock observation data: Acquire multi-source surrounding rock observation data under multiple consecutive excavation cycles of the tunnel. The multi-source surrounding rock observation data includes at least tunnel face image data, three-dimensional point cloud data, geological logging data, seepage observation data, rockfall area record data, and deformation monitoring data. The multi-source surrounding rock observation data is uniformly numbered, time-aligned, spatially registered, attribute-normalized, and structured-coded to form standardized input data.

[0012] S2. Extraction of surrounding rock structure objects and construction of graph nodes: Based on the standardized input data, extract joints, fissures, bedding, seepage points, spalling areas and abnormal deformation units, and encode the extraction results into graph nodes. Each graph node includes at least the object category, the excavation cycle to which it belongs, the spatial location, the geometric features, the attribute features, the quality evaluation value and the initial risk value.

[0013] S3. Relationship edge construction and single-cycle structure graph generation: Based on the spatial adjacency relationship, same group joint relationship, geometric continuity relationship, local co-occurrence relationship and risk propagation relationship between graph nodes, relationship edges are constructed to generate the single-cycle structure graph corresponding to the current excavation cycle;

[0014] S4. Cross-cycle node matching and graph memory network update: Perform cross-cycle node matching between the single-cycle structure graph corresponding to the current excavation cycle and the graph memory network corresponding to the historical excavation cycles, establish inheritance, splitting or merging relationships between the current node and the historical node, and update the node state and relationship edge state to form the updated graph memory network.

[0015] S5. Cross-cycle correlation analysis and typical pattern recognition: Based on the updated graph memory network, cross-cycle correlation analysis is performed to identify typical patterns such as cross-cycle extension of the same structural surface, transformation of cracks into spalling areas, expansion of seepage points into water-rich zones, and enhanced coupling between structural anomalies and monitoring anomalies, and to extract the anomaly expansion path.

[0016] S6. Accumulated Risk Assessment and Determination of Key Areas of Concern: Based on the current risk of nodes, historical manifestation intensity, abnormal evolution intensity, and path or pattern propagation contribution, the cumulative risk of nodes, relationship edge propagation risk, and regional risk are assessed, and key areas of concern are determined.

[0017] S7. Result Output and Graph Memory Continuous Incremental Maintenance: Output the structural surface continuity analysis results, abnormal expansion paths, risk accumulation levels, and key areas of concern. Write the new nodes, update relationships, state changes, risk results, and typical pattern results of the current excavation cycle into the graph memory network for continuous use in subsequent excavation cycles.

[0018] Based on the above technical solution, the present invention can be further improved as follows.

[0019] Furthermore, in step S1, the multi-source surrounding rock observation data are collected using the excavation cycle number as the basic organizational unit; the time alignment includes interpolating and synchronizing the continuous monitoring data using the end time of the current excavation cycle or the end time of the face clearing as the reference time; the spatial registration includes mapping the image coordinates, point cloud coordinates, equipment local coordinates, and monitoring point coordinates to a unified local engineering coordinate system of the face; the structured coding includes recording the excavation cycle number, object category, acquisition time, spatial location, geometric features, anomaly intensity, data source, confidence level, and quality identifier for each observation object.

[0020] Furthermore, in step S2, the surrounding rock structure objects are divided into three categories: linear structure objects, regional anomaly objects, and monitoring response objects. The linear structure objects include joints, fissures, and bedding; the regional anomaly objects include seepage points and spalling areas; and the monitoring response objects include deformation anomaly units. The directional and geometric features of the linear structure objects, the center position and regional range features of the regional anomaly objects, and the anomaly amplitude and rate of change features of the monitoring response objects are extracted to generate the graph nodes.

[0021] Furthermore, in step S3, the construction of the relation edges includes:

[0022] Perform candidate adjacency filtering on the graph nodes in the current excavation cycle;

[0023] Spatial adjacency edges are established based on the spatial distance between nodes;

[0024] Establish joint edges in the same group based on directional and attribute similarity between linear structure objects;

[0025] Geometric continuation edges are established based on directional consistency, endpoint distance, projection connectivity, and intermediate blocking conditions;

[0026] Establish local co-occurrence edges based on the synchronous appearance of different types of abnormal objects within the same local analysis unit;

[0027] A directed risk propagation edge is established based on the preset object category propagation rules, the anomaly strength of the source node, and the degree of influence of the neighborhood;

[0028] The set of graph nodes and the set of relation edges are then organized into the single-cycle structure graph.

[0029] Furthermore, in step S4, the cross-loop node matching includes:

[0030] Based on object category consistency, spatial proximity, directional compatibility, and structural semantic rationality, candidate historical nodes are selected.

[0031] The matching similarity between the current node and candidate historical nodes is calculated by combining location similarity, geometric feature similarity, attribute feature similarity, and neighborhood relationship similarity.

[0032] When the matching similarity is greater than the preset matching threshold and reaches the maximum value among the candidate objects, it is determined that the current node and the corresponding historical node constitute an inheritance matching relationship;

[0033] If one historical node corresponds to multiple current nodes, it is determined to be displayed as a split; if multiple historical nodes correspond to one current node, it is determined to be displayed as a merge.

[0034] The historical node status is marked as one of the following: active, not yet displayed, or terminated.

[0035] Furthermore, in step S5, the cross-cyclic association analysis includes:

[0036] Continuity analysis is performed on the node sequences corresponding to the same memory identifier to identify cross-cycle extension trajectories of the same structural plane;

[0037] The relationship changes of different types of nodes in multiple consecutive excavation cycles are tracked to identify local transformation chains between crack propagation, enhanced water seepage, spalling formation and deformation anomalies;

[0038] Path search is performed on directed propagation edges and multi-type relation edges in graph memory networks to identify anomalous expansion paths and risk propagation paths;

[0039] The subgraph structure is matched, the path features are clustered, or the rule templates are compared to form typical pattern recognition results;

[0040] Among them, the typical modes include at least the cross-cycle extension mode of the same structural plane, the transformation mode of cracks into slab areas, the expansion mode of seepage points into water-rich zones, the coupling mode of structural anomalies, water anomalies, and deformation anomalies, and the mode of repeated manifestation of local anomalies.

[0041] Furthermore, in steps S6 and S7, the determination of the key concern area includes: aggregating the node risk value and the relationship edge propagation risk value within the local window of the tunnel face, regular grid, spatial clustering unit, or key structural partition to obtain a comprehensive regional risk value; determining the key concern area based on the comprehensive regional risk value, the number of high-risk nodes within the area, the number of high-propagation risk edges, the abnormal expansion path, the typical mode, and the enhancement trend of multiple consecutive excavation cycles; and spatially extrapolating the key concern area along the extension direction of the linear structural object, the direction of the seepage propagation path, or the direction of the block fall expansion.

[0042] The output results and continuous incremental maintenance of the graph memory include: outputting the structural surface continuity analysis results, abnormal expansion paths, risk accumulation levels and key areas of concern, and writing newly formed nodes, relation edges, state changes, risk levels and typical patterns in the current excavation cycle into the graph memory network.

[0043] A graph memory-based cross-cycle correlation analysis system for tunnel surrounding rock structural surfaces includes:

[0044] The data access module is used to access face image data, 3D point cloud data, geological logging data, seepage observation data, rockfall area record data, and deformation monitoring data from multiple consecutive excavation cycles.

[0045] The data alignment and standardization module, connected to the data access module, is used to perform unified numbering, time alignment, spatial registration, attribute normalization, and structured coding on the multi-source surrounding rock observation data.

[0046] The node construction module, connected to the data alignment and standardization module, is used to extract joints, cracks, bedding, seepage points, spalling areas, and abnormal deformation elements, and encode them as graph nodes.

[0047] The relationship modeling module, connected to the node construction module, is used to construct relationship edges and generate a single-loop structure graph based on spatial adjacency, same-group joint relationship, geometric continuity relationship, local co-occurrence relationship and risk propagation relationship.

[0048] The graph memory update module, connected to the relationship modeling module, is used to match and merge the single-cycle structure graph corresponding to the current excavation cycle with the historical graph memory network, update the node state and relationship edge state, and form an updated graph memory network.

[0049] The cross-cycle correlation analysis module, connected to the graph memory update module, is used to identify typical patterns such as cross-cycle extension of the same structural surface, transformation of cracks into spalling areas, expansion of seepage points into water-rich zones, and enhanced coupling between structural anomalies and monitoring anomalies, and to extract the anomaly expansion path.

[0050] The risk assessment module, connected to the cross-cycle correlation analysis module, is used to cumulatively assess node risk, relationship edge propagation risk, and regional risk, and to determine key areas of concern.

[0051] The results output module, connected to the risk assessment module, is used to output the structural continuity analysis results, abnormal expansion paths, risk accumulation levels, and key areas of concern. It also writes the incremental data of newly added nodes, update relationships, state changes, risk results, and typical pattern results corresponding to the current excavation cycle into the graph memory network.

[0052] Furthermore, the graph memory update module includes: a candidate filtering unit, a cross-loop matching unit, a state update unit, a relationship inheritance unit, and a historical version maintenance unit;

[0053] The candidate filtering unit is used to filter candidate objects corresponding to the current node from historical memory nodes. The cross-cycle matching unit is used to perform similarity matching between the current node and historical nodes by comprehensively considering spatial location, geometric features, attribute changes and neighborhood structure. The state update unit is used to inherit the historical identifier of successfully matched nodes and update the state vector, number of appearances, duration and risk evolution fields. The relation inheritance unit is used to align, incrementally update or deactivate relation edges in the current cycle with historical edges. The historical version maintenance unit is used to save the graph state snapshot and key event log after each excavation cycle update.

[0054] Furthermore, the result output module includes: a graphical display unit, a structured result output unit, an interface service unit, and a historical archiving unit;

[0055] The graphical display unit is used to display the analysis results in a multi-level display mode of object layer, relationship layer, path layer and region layer. The structured result output unit is used to output standardized results by node, region, path, pattern and excavation cycle index. The interface service unit is used to provide result call interface to construction management platform, early warning system, digital twin system or mobile terminal. The historical archiving unit is used to save graph memory network snapshots, risk results, typical pattern results and key event logs, and supports subsequent excavation cycles to continue to perform incremental analysis on the basis of existing history.

[0056] Compared with the prior art, the technical solution of this application has the following beneficial technical effects:

[0057] 1. This invention constructs a cross-cycle graph memory network to uniformly associate and continuously store the manifestation results of objects such as joints, fissures, bedding, seepage points, spalling areas, and deformation anomalies in different excavation cycles. This breaks through the isolated analysis method of existing technologies that are mainly based on single cycles, single sections, and single moments. It can continuously record the manifestation process, evolution trajectory, and state changes of the surrounding rock structure, thereby improving the integrity, stability, and temporal consistency of the identification of the continuity of the surrounding rock structure.

[0058] 2. By establishing multiple types of relation edges, such as spatial adjacency, same group joints, geometric continuation, local co-occurrence, and risk propagation, this invention achieves a unified expression of complex relationships between surrounding rock structures. It can more accurately reveal the intrinsic connections between joint extension, crack propagation, water seepage propagation, rockfall formation, and monitoring anomaly responses, thereby enhancing the structured expression, logical interpretation, and engineering readability of surrounding rock anomaly identification results.

[0059] 3. This invention effectively tracks the continuous manifestation, local splitting, reconvergence, range expansion, and anomaly enhancement processes of the same structural face in different excavation cycles through cross-cycle node matching, historical state inheritance, and incremental relationship update mechanisms. This avoids the problems of historical information fragmentation, repeated judgment, or interruption of evolution chain caused by face advancement, data dispersion, or local occlusion in the prior art, and improves the continuity, accuracy, and reliability of cross-cycle identification results.

[0060] 4. This invention can identify typical patterns such as cross-cycle extension of the same structural surface, transformation of cracks into spalling areas, expansion of seepage points into water-rich zones, and enhanced coupling between structural anomalies and monitoring anomalies. It elevates the originally scattered and fragmented identification results into patterned results with evolutionary semantics, path semantics, and risk semantics, which is conducive to the earlier detection of signs of surrounding rock instability, abnormal development trends, and local risk evolution chains, thereby improving the ability to make advanced identifications and risk predictions during construction.

[0061] 5. This invention conducts a cumulative risk assessment by comprehensively considering node risk, relationship propagation risk, historical manifestation intensity, anomaly propagation path, local coupling degree, and the credibility of typical patterns. This enables dynamic quantification, continuous updating, and graded determination of surrounding rock risk. Compared to static judgments based solely on current cyclical local observation results, this invention more accurately reflects the persistence, accumulation, expansion, and propagation of surrounding rock anomalies, thereby improving the accuracy, relevance, and engineering applicability of risk assessment results.

[0062] 6. This invention can output structural continuity analysis results, abnormal expansion paths, risk accumulation levels, and key areas of concern. It also supports continuous incremental maintenance and historical process tracing of graph memory networks, providing continuous, traceable, and updatable data support and analysis basis for subsequent monitoring encryption, local verification, support optimization, risk warning, and construction organization adjustment. It has strong engineering practical value, field application value, and prospects for promotion and application. Attached Figure Description

[0063] Figure 1 The overall flowchart provided for embodiments of the present invention.

[0064] Figure 2 A flowchart illustrating the acquisition and unified representation of multi-source surrounding rock observation data provided in this embodiment of the invention;

[0065] Figure 3 This is a flowchart of the extraction of surrounding rock structure objects and construction of graph nodes provided in an embodiment of the present invention;

[0066] Figure 4 This is a flowchart of the relationship edge construction and single-loop structure graph generation provided in an embodiment of the present invention;

[0067] Figure 5 A flowchart for cross-cycle node matching, graph memory network update, association analysis, risk assessment, and result output provided in an embodiment of the present invention;

[0068] Figure 6 This is a schematic diagram of a single-loop structure provided in an embodiment of the present invention;

[0069] Figure 7 This is a schematic diagram of the system overview interface provided in an embodiment of the present invention;

[0070] Figure 8 This is a schematic diagram of the system correlation analysis interface provided in an embodiment of the present invention;

[0071] Figure 9 This is a statistical chart showing the changes in the number of nodes and the number of relational edges under continuous excavation cycles, provided in an embodiment of the present invention.

[0072] Figure 10 This is a statistical graph showing the distribution of cross-loop node matching similarity in an embodiment of the present invention.

[0073] Figure 11 This is a statistical chart showing the frequency of occurrence of typical patterns provided in embodiments of the present invention;

[0074] Figure 12 Statistical charts of node-level, edge-level, and region-level risk assessment results provided for embodiments of the present invention;

[0075] Figure 13 The system computing performance and resource consumption statistics provided for embodiments of the present invention are shown in the figure. Detailed Implementation

[0076] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings, which illustrate embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of this application will be thorough and complete.

[0077] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0078] Example 1:

[0079] refer to Figures 1-13 In this embodiment, a construction section of a mountain railway tunnel is taken as the object, and multiple consecutive excavation cycles are selected as analysis units. After each excavation cycle is completed, multi-source data collection and structural object identification are performed on the tunnel face and its adjacent areas, and the incremental analysis results of each cycle are written into the graph memory network.

[0080] Based on this, a cross-cycle correlation analysis method for tunnel surrounding rock structural surfaces based on graph memory is implemented. The specific implementation steps include:

[0081] I. Acquisition and Unified Representation of Multi-Source Surrounding Rock Observation Data

[0082] In this embodiment, the acquisition and unified representation of multi-source surrounding rock observation data refers to the collection of images, point clouds, geological logs, seepage records, rockfall records, and monitoring data of the surrounding rock at the tunnel face under multiple consecutive excavation cycles during tunnel construction. All types of data are then uniformly numbered, time-aligned, spatially registered, attribute-normalized, and structured-coded to form a standardized input dataset suitable for subsequent graph node construction and cross-cycle correlation analysis.

[0083] Specifically, with the first Each excavation cycle is used as the basic analysis unit. The multi-source observation data corresponding to this cycle are organized into a set of cycle-level data, which includes at least: face image data, 3D point cloud data, geological sketches and surrounding rock logging data, seepage anomaly observation data, data recorded in rockfall areas or locally fractured areas, and monitoring data such as convergence deformation, crown settlement, and surrounding rock displacement. Preferably, the first... The original data set for each excavation cycle can be represented as: ;

[0084] in, Indicates the first Image data of the excavation face during one excavation cycle. Represents point cloud data. This refers to geological logging or geological sketching data. This represents the data from water seepage observations. This indicates the data recorded in the block drop area or the local instability area. This indicates monitoring data.

[0085] In one embodiment, the face image data is acquired by an industrial camera, handheld terminal, or mobile inspection equipment to reflect the surface texture, fracture boundaries, wetted areas, and spalling marks of the face; the 3D point cloud data is acquired by a laser scanner, structured light equipment, or photogrammetry system to describe the geometric morphology, local concave-convex features, and spatial distribution information of structural surfaces of the face; the geological logging data is entered by on-site geologists through a mobile terminal or geological recording system, including the surrounding rock type, joint group, bedding distribution, weak interlayers, fault fracture zones, and related textual descriptions; the seepage observation data records the seepage location, seepage intensity, distribution range, and duration; the spalling record data is used to describe local block spalling, loosened areas, and surface damage areas; and the monitoring data includes convergence values, settlement values, displacement, rates, and their temporal changes.

[0086] Because the data sources, sampling frequencies, and recording formats differ, a unified representation and processing method is required. First, each source data is assigned a unified excavation cycle number, collection timestamp, and data source identifier. For data with inconsistent sampling times within the same excavation cycle, synchronization is achieved using a reference time alignment method. Preferably, the end time of the cycle or the completion time of face cleaning for that cycle is used as the reference time. For continuous monitoring data, linear interpolation is used to obtain observations at a uniform time. The expression can be written as:

[0087] ;

[0088] in, and Representing reference time Two consecutive sampling times, and This represents the original observation value at the corresponding time. Indicates alignment to the reference time. The unified observation values; in this way, monitoring curves, seepage observations and local records can be mapped to the same cycle time for analysis.

[0089] Furthermore, to eliminate inconsistencies between image coordinates, point cloud coordinates, device local coordinates, and engineering coordinates, spatial unification is performed on the multi-source data. Specifically, image pixel positions, depth values, or point cloud point coordinates can be mapped to the local engineering coordinate system of the tunnel face through a coordinate transformation matrix. Preferably, the spatial unification process can be expressed as: ;

[0090] in, Represents the pixel coordinates of the image. This represents depth information or equivalent distance information. Represents the mapped engineering coordinates. It represents the spatial transformation matrix composed of equipment calibration parameters, attitude parameters, and coordinate system transformation parameters; through this mapping, cracks, seepage points, rockfall boundaries, point cloud geometric features, and monitoring point positions in the face image can be uniformly expressed under the same spatial reference frame.

[0091] After completing time alignment and spatial registration, various attributes need to be standardized. Since attributes such as crack length, seepage area, spalling area, displacement, and growth rate have different dimensions and large numerical ranges, continuous attributes are normalized to ensure the stability of subsequent node modeling and correlation analysis. Preferably, range normalization can be used. ;

[0092] in, Indicates the attribute value to be processed. and These represent the minimum and maximum values ​​of the attribute in the current data batch or historical statistical sample, respectively. To prevent tiny positive numbers with a denominator of zero, This represents the normalized attribute value; after normalization, attributes from different sources and with different dimensions can participate in subsequent graph node attribute calculations and risk assessments in a unified numerical space.

[0093] To facilitate subsequent graph modeling, this embodiment also organizes the multi-source observation results into unified data entries. For any observation object, at least the following fields are recorded: excavation cycle number, object category, acquisition time, spatial location, geometric features, anomaly intensity, data source, confidence level, and quality identifier. Preferably, the unified expression vector of a single observation object can be written as:

[0094] ;

[0095] in, Indicates the first Each excavation cycle number, Indicates the object category identifier. Represents a spatial location vector. Represents geometric eigenvectors. Represents the attribute feature vector. This indicates the confidence level of the object's identification or the data quality evaluation value.

[0096] In a preferred embodiment, the object categories include at least joints, fissures, bedding, seepage points, spalling areas, and deformation anomaly units; the geometric features include at least one or more of length, width, area, boundary range, orientation angle, or morphological description; the attribute features include at least one or more of seepage intensity, spalling severity, displacement increment, growth rate, and historical manifestation markers. For text-based geological logging information, it can be first converted into structured attributes through field extraction or rule parsing, and then written into a unified expression vector.

[0097] In addition, data quality can be screened during the unified expression stage. Data entries with blurry images, severe point cloud defects, abnormal jumps in monitoring values, or incomplete manual records are assigned a low-quality label or low confidence level so that their weight can be reduced or they can be removed in the subsequent node construction and relationship modeling stages. For observation entries with missing data but still of reference value, historical statistical values, nearby measurement point interpolation, or rule-based completion methods can be used for correction, and the completion label is retained to ensure the traceability of the subsequent analysis process.

[0098] Through the above processing, the first Multi-source surrounding rock observation data from images, point clouds, logging, seepage, rockfall, and monitoring systems during each excavation cycle are uniformly organized into standardized input data with a unified time base, unified spatial reference, unified attribute scale, and unified structural format. This standardized input data can directly support subsequent processes such as surrounding rock structure object extraction, graph node construction, relation edge generation, cross-cycle matching, and graph memory update, thus providing a reliable data foundation for the entire graph memory-based cross-cycle association analysis method.

[0099] II. Extraction of Surrounding Rock Structure Objects and Construction of Graph Nodes

[0100] After completing the unified representation of multi-source surrounding rock observation data, this step further extracts surrounding rock structural objects from the standardized input data and encodes the identified joints, fissures, bedding, seepage points, spalling areas, and deformation anomaly units into graph structure nodes. The core of this step is: on the one hand, converting surrounding rock information from different sources, with different shapes, and different scales into a unified object-level representation; on the other hand, generating node data with spatial, structural, and state attributes for subsequent relationship modeling, cross-cycle matching, and risk evolution analysis.

[0101] In one embodiment, the extraction of surrounding rock structure objects does not rely solely on a single data source, but rather on joint identification based on images, point clouds, geological logging, inspection records, and monitoring results. For structural surface objects such as joints, fissures, and bedding planes, a comprehensive judgment is preferably made by combining texture edge information in the face image, local normal variation information in the point cloud, and phylogenetic records in the geological logging. For seepage point objects, joint confirmation is preferably made using wetted boundaries, highly reflective areas, color anomaly areas in the image, and on-site seepage records. For slab drop areas, identification is preferably made based on block boundaries in the image, local depression areas in the point cloud, surface gap areas, and manually marked records. For deformation anomaly units, extraction is performed based on incremental changes, rate changes, and spatial neighborhood consistency of monitoring point data.

[0102] To ensure that different types of objects can be uniformly included in the graphical model, this embodiment divides the surrounding rock structure objects into three categories: linear structure objects, regional anomaly objects, and monitoring response objects. Among them, joints, fissures, and bedding belong to linear structure objects, which are usually characterized by slender features with directionality, ductility, and systematization attributes; seepage points and spalling areas belong to regional anomaly objects, which are usually characterized by local clusters, relatively closed boundaries, or measurable range of anomaly areas; deformation anomaly units belong to monitoring response objects, which usually correspond to the numerical anomaly response of discrete measuring points or local measuring areas. Through this classification method, the geological structure expression and construction risk expression can be taken into account in the unified graphical structure.

[0103] In a preferred embodiment, object extraction first constructs candidate sets according to object categories, using the first... Taking an excavation cycle as an example, let the set of structured objects extracted in this cycle be: ;

[0104] in, Indicates the first The first one identified during the excavation cycle A surrounding rock structure object, This indicates the total number of objects in the current loop.

[0105] Furthermore, each object is assigned to a subset of joint objects, a subset of fracture objects, a subset of bedding objects, a subset of seepage objects, a subset of spalling objects, or a subset of abnormal deformation objects according to its category. This object-level organization method is beneficial for subsequent execution of classification attribute extraction and differential value assignment.

[0106] For structural surface objects such as joints, fissures, and bedding, they are preferably characterized from two levels: geometric shape and directional features. Geometric shape includes at least length, width, area, boundary extent, and connectivity. Directional features include at least orientation, dip, local principal direction, or extension axis. For structural objects that appear as linear boundaries in an image, their centerline and boundary can be obtained through edge tracing, connected component extraction, and principal axis fitting. For structural objects that appear as local planes or abrupt changes in direction in a point cloud, their normal features and local plane parameters can be extracted through neighborhood fitting or normal clustering. Preferably, the principal direction angle of a structural surface object can be calculated by fitting its boundary point set or centerline point set, expressed as: ;

[0107] in, and These represent the coordinates of the two endpoints of the object's principal axis or the two endpoints of the fitted line segment, respectively. Indicates the first The object in the first The main direction angle in each cycle To prevent tiny positive numbers with a denominator of zero.

[0108] This method can provide a directional basis for joint group division, continuity relationship identification, and cross-cycle matching.

[0109] For anomalous objects such as seepage points and spalling areas, it is preferable to extract attributes such as their center location, area, boundary compactness, anomaly intensity, and neighborhood distribution density. In addition to spatial location, seepage point objects can also record attributes such as seepage intensity level, seepage range, and whether it extends in a banded pattern; spalling area objects can also record, in addition to area, boundary fragmentation degree, number of adjacent cracks, and severity of local damage. Preferably, the equivalent scale of the regional objects can be represented using the area equivalent radius method: ;

[0110] in, Indicates the first The abnormal object in the region is in the first The area in each cycle, This represents the corresponding equivalent scale parameter. This parameter can be used for subsequent scale comparisons between regional objects, neighborhood searches, and risk propagation range analysis.

[0111] For deformation anomaly units, this embodiment preferably uses monitoring points or monitoring zones as basic object carriers, and generates object attributes based on their time-period increments, change rates, and anomaly discrimination results. For the abnormal response of a certain monitoring point in the current cycle, it can be measured according to the degree of deviation of its observed value from the baseline value, the previous cycle value, or historical statistical levels. Preferably, the deformation anomaly intensity can be expressed as:

[0112] ;

[0113] in, Indicates the first The monitoring object was in the first The current observation value in each loop, This indicates its historical mean or reference baseline. This indicates the corresponding historical fluctuation scale or standard deviation. This represents the standardized anomaly intensity. This expression method standardizes the anomaly scale for different types of monitoring quantities, providing a basis for subsequent correlation between monitoring anomaly nodes and structural anomaly nodes.

[0114] After object extraction, the objects need to be further encoded into graph nodes. Unlike simply storing detection results, the nodes in this invention are not only containers for identification results, but also basic computational units for subsequent graph relationship modeling, cross-cycle inheritance, and risk propagation analysis. Therefore, in addition to retaining the geometric and attribute information of the object itself, each node preferably also includes category labels, cycle labels, status labels, and quality labels. Specifically, a node includes at least the following fields: unique node identifier, object category, excavation cycle to which it belongs, spatial location, geometric feature vector, attribute feature vector, data quality evaluation value, and initial risk value. For objects that need to be continuously tracked across cycles, the node can also reserve historical inheritance identifier fields and matching candidate fields for direct retrieval in subsequent graph memory update steps.

[0115] In a preferred embodiment, the graph node can be represented as follows: ;

[0116] in, Indicates the unique identifier of the node. Indicates the object category identifier. Indicates spatial location parameters, Represents geometric eigenvectors. Represents the attribute feature vector. This indicates a quality or confidence level rating. This represents the initial risk value. In this way, objects of different categories can be uniformly written into the node collection while retaining their differentiated attributes.

[0117] Preferably, the initial risk value is not a fixed constant, but is generated comprehensively based on the object category and key attributes. For structural surfaces such as joints, fissures, and bedding, values ​​can be assigned based on factors such as length, aperture, unfavorable orientation, and proximity to existing anomalous areas; for seepage points, values ​​can be assigned based on factors such as seepage intensity, persistence, and expansion trend; for spalling areas, values ​​can be assigned based on area, boundary irregularity, and nearby fissure density; for deformable anomalous objects, values ​​can be assigned based on anomalous intensity, growth rate, and spatial linkage. Preferably, the initial risk value of a node can be expressed as: ;

[0118] in, Indicates the first The node at the th The first cycle Each risk-related attribute Indicates the corresponding attribute weight. This indicates the number of attributes involved in risk assignment. This weighting method allows for the initial reflection of risk differences among different objects during the node generation stage, providing a foundation for subsequent risk propagation and accumulation analysis.

[0119] Furthermore, this embodiment preferably introduces a node validity screening mechanism during the node construction process. For unstable objects caused by low-quality images, local point cloud defects, incomplete records, or monitoring noise, if their confidence level or quality evaluation value is lower than a preset threshold, they can be marked as candidate nodes or weak nodes, and will not be directly included in the main graph node set; for objects supported by multi-source information, a higher confidence level is assigned. This mechanism helps reduce the interference of misidentified objects on subsequent relationship modeling and cross-loop matching.

[0120] After completing the node encoding, the first The set of nodes for an excavation cycle can be represented as: ;

[0121] This set of nodes forms the basis of the current cyclic surrounding rock structure diagram. Subsequent steps will further construct the spatial adjacency, group relationship, continuity relationship and risk propagation relationship between nodes, and generate a single-cycle structure diagram and a cross-cycle diagram memory network.

[0122] It should be noted that the object extraction method in this step is not limited to a specific identification algorithm, nor is it limited to a specific device data source. Any method that can convert surrounding rock structure objects under different cycles into node forms with unified semantic fields, unified spatial representation, and unified state representation can be applied to the technical solution of this invention. Through the implementation of this step, the originally scattered structural observation results in the tunnel face are transformed into a set of nodes with clear structure, complete attributes, and usable for time-series tracing and graph association analysis, thus providing a reliable foundation for subsequent relationship modeling, cross-cycle matching, and risk evolution identification.

[0123] III. Construction of Relational Edges and Generation of Single-Loop Structure Graphs

[0124] After extracting the surrounding rock structure objects and constructing graph nodes under the current excavation cycle, this step further identifies multiple types of relationships between nodes based on the nodes, and constructs graph edges according to different relationship types, thereby generating the first... The single-cycle structure diagram corresponding to each excavation cycle. Unlike methods that only save discrete detection results, this invention explicitly constructs the relationship edges between nodes, so that joints, cracks, bedding, seepage points, spalling areas, and abnormal deformation units no longer exist as isolated objects, but are organized into a graph network with internal structural connections and local evolutionary semantics, providing a foundation for subsequent cross-cycle matching, evolutionary path identification, and risk propagation analysis.

[0125] In one embodiment, the construction of relation edges is not an indiscriminate connection of all nodes. Instead, candidate adjacency screening is first performed on the set of nodes in the current loop, and then edge types and weights are generated according to different relation determination criteria. That is, graph edges are only established between two nodes when they meet the corresponding constraints in spatial location, structural attributes, directional characteristics, local co-occurrence states, or risk impact mechanisms. This approach avoids overly dense graph structures while enhancing the physical rationality and engineering interpretability of relation representations.

[0126] Preferably, the first The set of candidate edges under one excavation cycle can be represented as: ;

[0127] In the actual construction process, not all candidate edges are directly assigned values. Instead, an initial screening is performed based on the distance between nodes, their category, and local neighborhood features. Node pairs with spatial distances exceeding a preset range, logically unrelated categories, or excessively low quality evaluation values ​​can be directly eliminated to reduce the size of the edge set to be judged.

[0128] 1. Spatial adjacency relationship construction

[0129] Spatial adjacency describes the geometric proximity of two surrounding rock structure objects in the current loop. This relationship is applicable to various object combinations, such as joints and fissures, fissures and seepage points, spalling areas and adjacent fractured structures, and monitoring anomalies and local anomaly areas. The role of spatial adjacency is to reflect the possible direct contact, short-range influence, or potential coupling between local anomaly objects.

[0130] In a preferred embodiment, the spatial distance between node pairs is first calculated based on node location parameters. If the nodes are point objects, Euclidean distance is used; if the nodes are linear objects or region objects, center distance, minimum boundary distance, or principal axis projection distance can be used as the metric. For any two nodes... and Its location distance can be written as: ;

[0131] in, and These represent the position parameters of the two nodes, respectively. If... If the distance is less than the adjacency threshold, the node pair is considered to meet the spatial adjacency candidate condition. Furthermore, the adjacency strength can be assigned a value based on the distance; the closer the distance, the higher the spatial adjacency edge weight.

[0132] Preferably, spatial adjacency relationships are not only considered in terms of absolute distance, but can also be modified by incorporating the object scale. For example, for large-area areas of landslides or large-scale areas of abnormal seepage, the adjacency influence range should be appropriately enlarged; for small-scale micro-cracks, the adjacency range is relatively reduced. By introducing object scale parameters, the adaptability of adjacency relationship determination can be improved.

[0133] 2. Construction of joint relationships within the same group

[0134] Joint group relationships are primarily used to characterize the consistency of system properties among multiple joints, fissures, or bedding objects. Since structural surfaces in tunnel faces often develop in groups, if multiple structural objects have similar directional characteristics, similar geometric shapes, and similar development trends, they are likely to belong to the same structural surface system. This relationship plays a crucial role in identifying local structural control features, tracking the distribution of similar structures in the tunnel face, and subsequent cross-cycle continuation analysis.

[0135] In one embodiment, for nodes belonging to the linear structure object category, features such as their principal orientation angle, length, opening, and apparent continuity are first extracted, and their grouping relationship is determined based on orientation similarity and attribute similarity. Preferably, the orientation difference between two structure surface nodes can be expressed as:

[0136] ;

[0137] when If the joints are less than a preset directional threshold and both structures meet similarity constraints in length, aperture, or surface features, a joint edge can be established between them. This edge represents the consistency of the two structural objects in their grouping, rather than simple spatial proximity. Therefore, even if the two structures are not spatially adjacent, a grouping relationship can still be established as long as the grouping criteria are met.

[0138] Preferably, for cases where joint logging records already exist on-site, the joint group number, joint occurrence range, and geological personnel's interpretation results in the logging records can be used as auxiliary constraints to improve the engineering credibility of the joint relationship construction.

[0139] 3. Construction of geometric continuation relationships

[0140] Geometric continuity describes whether two structural objects have geometrical continuity, extension, or projective connectivity in the current working face. This relationship differs from co-group joint relations, which emphasize group consistency, while geometric continuity emphasizes whether specific objects may form part of the same extended structure. This relationship is particularly useful for identifying fractures, joints, and bedding objects that are partially occluded, discontinuously visible, or segmentally exposed.

[0141] In a preferred embodiment, the system, for linear structural objects or hybrid linear and region objects, comprehensively considers directional consistency, endpoint distance, projection connectivity, and intermediate blocking conditions to determine whether to establish a geometric continuation edge. For two candidate structural objects, if their main directions are close, their ends face each other, the distance between them is small, and there are no obvious discontinuities between them, then it is determined that there is a possibility of geometric continuation between them. Preferably, the geometric continuation score can be written as:

[0142] ;

[0143] in: , , These are the weighting coefficients; , For attenuation parameters; This indicates the evaluation value for the degree of endpoint docking, boundary connectivity, or continuity of intermediate areas. If If the value exceeds a preset threshold, a geometric continuation edge is established between the corresponding nodes.

[0144] In one implementation, geometric continuity can also be used to connect the boundary directions of crack nodes and spalled area nodes. For example, when the end of a crack points to the local spalled boundary and the boundary directions of both are consistent, a weak continuity edge can be established to characterize the control relationship between crack development and the local spalled boundary.

[0145] 4. Construction of local co-occurrence relationships

[0146] Local co-occurrence relationships are used to describe the synchronous or accompanying occurrence of different types of anomalous objects within the same local area. Compared to spatial adjacency relationships, local co-occurrence relationships place greater emphasis on the coordinated occurrence of multiple objects within the same local unit. Examples include the simultaneous occurrence of densely fractured areas and seepage points, the co-occurrence of areas with enhanced seepage and spalling, and the synchronous activity of local structural anomalies and deformation anomalies. This relationship reflects the complex anomaly characteristics of the local surrounding rock condition.

[0147] In one embodiment, the system divides the working face into several local analysis units, or constructs local neighborhood windows centered on nodes. When two or more nodes are located in the same local analysis unit and appear simultaneously in the current loop, local co-occurrence edges can be established between the corresponding nodes. For co-occurrence among multiple types of nodes, different edge weights can be assigned based on co-occurrence frequency, local density, and object category combination type.

[0148] Preferably, the local co-occurrence relationship is not limited to a binary determination of "whether they occur simultaneously," but can also consider the co-occurrence intensity. For example, if a region simultaneously exhibits high-intensity seepage, concentrated fissure development, and a sudden increase in monitoring values, its co-occurrence weight can be higher than that of a region with only slight seepage and isolated fissures. In this way, the single-cycle diagram can preserve the structural semantics of local composite risks.

[0149] 5. Establishing Risk Transmission Relationships

[0150] Risk propagation relationships are a crucial type of graph edge in this invention, used to express the inducing, enhancing, or transmitting effects of one type of structural anomaly on another. This relationship differs from simple spatial and geometric relationships; its focus is on characterizing the risk evolution logic between surrounding rock anomalies. For example, fracture propagation may lead to deterioration of block boundary conditions, thereby increasing the probability of rockfall; enhanced permeability may cause softening of the surrounding rock, thus promoting local deformation; dense joint development may reduce the integrity of the surrounding rock, making adjacent areas more prone to rockfall and increased convergence.

[0151] In a preferred embodiment, the risk propagation relationship is constructed using a combination of "category rule constraints + attribute strength constraints + neighborhood influence constraints". First, propagable object category pairs are predefined based on engineering mechanisms, such as "crack → spalling", "seepage → deformation", "dense joint zone → seepage expansion", "spalling area → further local instability", etc. Second, the source node is required to meet certain conditions in terms of abnormal intensity, scale, or growth trend. Third, the source node and the target node are required to have a basic association in their spatial neighborhood or local structure. When the above conditions are met, a directed risk propagation edge is established between the two nodes.

[0152] Preferably, the weight of the risk propagation edge can be expressed as:

[0153] ;

[0154] in, Indicates the matching degree of the node category propagation rule. Indicates the anomaly strength or risk level of the source node. This indicates the degree of influence in the neighborhood or the vulnerability assessment value of the target node. , , For weighting coefficients. When When the propagation threshold is exceeded, and Establish a directed risk propagation boundary between them.

[0155] In a specific scenario, when the length and aperture of a crack node increase, and its end is adjacent to a newly emerging spalling zone node, the system can establish a risk propagation edge from the crack node to the spalling zone node. Similarly, when the strength of a seepage point node increases, and the abnormal strength of adjacent monitoring points increases synchronously, the system can establish a propagation edge from the seepage node to the deformation anomaly node. Through this construction method, the single-loop structure diagram can not only express "what the object is and where it is," but also "which type of anomaly is affecting which type of anomaly."

[0156] 6. Edge attribute organization and single-loop structure graph generation

[0157] After determining the various types of relationships, the system assigns edge attributes to each established edge. Preferably, the edge attributes include at least a unique edge identifier, start and end node identifiers, edge type, edge weight, directional marker, establishment basis, and quality evaluation value. For undirected relationships, such as spatial adjacency or joint relationships, undirected edges can be used; for risk propagation relationships with significant influence directions, directed edges are preferred. For the same pair of nodes, multiple edge types are allowed to coexist to preserve richer relationship semantics.

[0158] Furthermore, the first The single-cycle structure diagram corresponding to one excavation cycle can be represented as follows:

[0159] ;

[0160] in, This represents the set of nodes in the current loop. This represents the set of relational edges in the current cycle. This single-cycle structure diagram reflects not only the surrounding rock structure state of the tunnel face at the current moment, but also the local connections, system consistency, geometric continuity characteristics, and risk impact links between different structural objects.

[0161] In a preferred embodiment, after generating the single-cycle structure graph, local simplification and structural correction processes can be performed on the graph. For example, weak edges that are obviously repetitive, unstable edges caused by low-quality nodes, or abnormal edges that obviously conflict with engineering constraints can be removed or their weights reduced; stable edges supported by multiple source data can have their weights increased or trusted markers added. The corrected single-cycle structure graph will serve as the direct input for subsequent cross-cycle node matching and graph memory updates.

[0162] 7. The Implementation Effect of this Step

[0163] Through this step, the previously scattered surrounding rock structure objects are organized into a single-cycle structure diagram with clear relational semantics. Compared to the traditional "object list" recording method, this invention can form a local correlation network between surrounding rock anomalies at the current excavation cycle level, preserving the attribute information of individual objects while explicitly expressing the spatial connections, group connections, continuity trends, and risk effects between objects. This eliminates the need for subsequent steps to infer the relationships between objects from scratch, allowing for direct cross-cycle inheritance, evolutionary pattern recognition, and risk accumulation analysis based on the structured diagram. This improves the accuracy and interpretability of surrounding rock structure continuity interpretation and local risk identification.

[0164] IV. Cross-cycle node matching and graph memory network update

[0165] After constructing the single-cycle structure diagram for the current excavation cycle, this step further matches the node set of the current cycle with the node sets of historical cycles to identify the continuous manifestation relationship of the same surrounding rock structure object in different excavation cycles, and updates the graph memory network based on this. Unlike the single-cycle structure diagram, which only reflects the state of the surrounding rock at the current moment, the graph memory network is used to record the manifestation, extension, enhancement, transformation, attenuation, and termination processes of the surrounding rock structure object during the continuous advancement of the tunnel face, thereby expanding the surrounding rock structure analysis from "current identification" to "cross-cycle evolution memory".

[0166] In one embodiment, let the first The graph memory network updated after each excavation cycle is as follows: , No. The single-cycle structure diagram corresponding to each excavation cycle is as follows: The purpose of this step is to... Nodes and edges in By mapping, inheriting, and merging historical objects in the graph, an updated graph memory network is obtained. .

[0167] 1. Candidate filtering for cross-loop matching objects

[0168] Since each node in the current loop may theoretically be similar to multiple nodes in the historical loop, this embodiment preferably performs candidate filtering first, rather than directly performing a full match on all historical nodes. Candidate filtering is mainly constrained by object category consistency, spatial proximity, directional compatibility, and structural semantic rationality, in order to narrow the matching range and reduce the probability of false matches.

[0169] Specifically, for any node in the current loop The system searches the historical memory network for nodes that meet the following conditions as candidate matching objects: First, the historical node and the current node belong to the same object category, or belong to the object category that allows transformation and association; Second, the last appearance position of the historical node and the position of the current node are within a preset spatial window; Third, for linear structure objects, the two satisfy compatibility constraints in terms of main direction, geometric scale, and local relationship pattern; Fourth, for regional anomaly objects and monitoring anomaly objects, the two are comparable in terms of center position, regional range, or measurement point mapping relationship.

[0170] Preferably, candidate screening can first construct a local candidate set based on the spatial search radius, and then perform secondary filtering according to category rules and attribute thresholds. This method can eliminate node pairs that lack the possibility of continuation in advance, thereby improving the efficiency of subsequent matching calculations.

[0171] 2. Node similarity calculation and matching determination

[0172] After forming the candidate set, the cross-cycle similarity between the current node and each candidate historical node is calculated. The similarity is not a single geometric distance, but a comprehensive consideration of factors such as spatial proximity, directional consistency, geometric scale continuity, attribute evolution rationality, and neighborhood structure similarity.

[0173] In a preferred embodiment, node matching similarity can be expressed as:

[0174] ;

[0175] in, This represents a historical node in a graph memory network. Indicates positional similarity. Represents geometric feature similarity. Indicates the similarity of attribute features. Indicates the similarity of neighborhood relationships. These are the corresponding weighting coefficients.

[0176] Among them, positional similarity is used to describe the spatial continuity of the current node and historical nodes in the face advancement coordinates; geometric feature similarity is used to describe the continuity of objects such as joints, fissures, and bedding in terms of length, scale, direction, and boundary morphology; attribute feature similarity is used to describe whether attributes such as seepage intensity, spalling severity, and anomaly level vary within a reasonable range; and neighborhood relationship similarity is used to compare whether the adjacent structures of two nodes in their respective graphs are consistent, such as whether there are the same set of joints, seepage anomalies, or monitoring anomalies in their vicinity.

[0177] Preferably, when the similarity between a candidate historical node and the current node is... If the score is greater than the matching threshold and reaches the maximum value among all candidate nodes, the current node is determined to have an inheritance matching relationship with that historical node. If multiple candidate nodes have similar scores, uniqueness constraints can be added, the most recently displayed object can be prioritized, or a manual verification flag can be introduced to avoid mismatches between one-to-many or many-to-one nodes.

[0178] 3. Node inheritance, splitting, and merging determination

[0179] In actual construction, the appearance of the same structural surface in different cycles is not always a simple one-to-one continuation relationship. Situations such as localized splitting, multiple segments re-converging, or abnormal expansion and merging of regions may also occur. Therefore, this embodiment, in addition to ordinary inheritance matching, preferably considers two types of scenarios: node splitting and node merging.

[0180] In the case of segmentation, when a historical node corresponds to multiple current nodes that are spatially close, oriented in the same direction, and connected by boundaries in the current loop, it can be determined that a historical object is being displayed in segments in the new loop. In this case, multiple current nodes share the same historical source identifier, but each records its local branch number and current state. This method is suitable for situations such as incomplete exposure of the tunnel face, partial occlusion, or discontinuous exposure of structural surfaces.

[0181] In the case of merging, when multiple historical nodes correspond to a single current node with a larger scope and stronger continuity in the current cycle, it can be determined that multiple historical fragments present a unified structural object in the new cycle. In this case, the current node can be marked as a merging node, and its mapping relationship with multiple historical source nodes can be preserved in the graph memory to ensure the integrity of the evolutionary trajectory.

[0182] Preferably, in the case of splitting or merging, the system records not only the inheritance relationship but also the relationship type identifier, so as to distinguish different evolution modes such as "continuous extension", "local discontinuity" and "structural convergence" during subsequent cross-cycle association analysis.

[0183] 4. Inheritance of historical states and updating node attributes

[0184] After the matching relationship is determined, the system performs node inheritance and attribute updates. For the current node that has successfully matched, it is no longer treated as a completely new object, but rather inherits the memory identifier of the historical node and writes the current observation value into its state sequence. The state sequence may include the display cycle number, position changes, geometric changes, attribute changes, and local relationship changes.

[0185] Preferably, for a node that has successfully inherited, its update status can be represented as: ;

[0186] in, This indicates the historical memory state of the node in the previous cycle. This represents the node state vector observed in the current loop. This indicates the updated memory state. This method preserves historical coefficients. By doing so, it is possible to absorb new information from the current cycle while preserving historical continuity, ensuring that node memory is neither entirely dependent on current observations nor loses historical evolutionary characteristics.

[0187] In one embodiment, if the current node exhibits characteristics such as increased scale, enhanced water seepage, higher anomaly level, expanded boundary, or strengthened local relationships compared to historical nodes, the corresponding increment is recorded in the node attributes; if the current node only shows a slight positional shift or minor fluctuation in its manifestation range, it is recorded as a normal continuation; if the anomaly intensity of the current node is significantly weakened or its manifestation boundary is significantly shrunk, it is marked as a decay state. Through this mechanism, the graph memory network can not only express "whether the object still exists," but also "how the object is changing."

[0188] 5. New node writes and retention of inactive nodes

[0189] For a node in the current loop, if no candidate matching the criteria is found in the historical memory network, it is determined to be a new node. New nodes typically correspond to newly exposed joints and fissures, newly appearing seepage points, newly formed slab areas, or monitoring points that show abnormal responses for the first time. For such nodes, the system generates a new memory identifier and writes it into the graph memory network as a new historical source.

[0190] Meanwhile, for certain nodes in the historical memory network, if no suitable successor is found in the current cycle, they should not be deleted immediately, but their historical records should be retained and their activity status updated. Preferably, historical nodes can have three states: "active," "not yet displayed," and "terminated." If a node successfully matches in the current cycle, it is in an active state; if a node is not displayed only in the current cycle but is geologically likely to continue appearing in subsequent cycles, it is marked as not yet displayed; if a node has not appeared for several consecutive cycles and does not meet the conditions for subsequent continuation, it can be marked as terminated. In this way, the loss of historical information due to local occlusion, missing data collection, or short-term non-display can be avoided.

[0191] 6. Update edges across loop relationships

[0192] In addition to the need for cross-cycle inheritance of nodes themselves, the edges between nodes also need to be updated synchronously. For spatial adjacency relationships, joint relationships, continuation relationships, and risk propagation relationships formed by inherited nodes in the current cycle, the system aligns them with the corresponding edges in the historical network and updates the duration, weight changes, and state changes of the edges. For newly established relationship edges in the current cycle, they are written into the graph memory network as new temporal relationships; for historical edges that do not reappear in the current cycle, their weights may be reduced or they may be marked as temporarily inactive, depending on the situation.

[0193] Preferably, the cross-cycle edge update not only records "whether it exists", but also "how many cycles it lasted", "how the intensity changed", and "whether it changed from weak to strong or from strong to weak". For example, if the same group relationship between two fractures persists in multiple consecutive cycles, the historical confidence of the relationship can be improved; if a risk propagation edge from a fracture to a drop block area is supported by observations in multiple consecutive cycles, the propagation path can be given higher weight in subsequent risk analysis.

[0194] 7. Construction and Organization Methods of Graph Memory Networks

[0195] After completing node inheritance, attribute updates, state annotations, and relation edge updates, we obtain the first... A graph memory network under a recurrence. Preferably, the graph memory network can be represented as: ;

[0196] in, This represents the updated set of memory nodes. This represents the updated set of memory edges. This represents the set of temporal state information associated with nodes and edges. The temporal state information includes at least one or more of the following: first appearance cycle, most recent appearance cycle, number of consecutive appearances, cumulative number of appearances, state change sequence, and relation evolution record.

[0197] In one embodiment, the graph memory network can be organized using a "current state layer + historical trajectory layer". The current state layer stores the state information of each object in the most recent loop, facilitating its rapid participation in the next loop matching; the historical trajectory layer stores the complete evolution record of an object from its first appearance to the current moment, facilitating subsequent path analysis, abnormal pattern recognition, and risk accumulation assessment. This two-layer organization method balances real-time update efficiency with the complete traceability of historical information.

[0198] 8. The Implementation Effect of this Step

[0199] Through this step, the single-cycle structure diagram in the current cycle is effectively incorporated into the historical evolution framework. Originally independent surrounding rock structures under different excavation cycles are linked together into a sequence of memory nodes with continuous manifestation relationships and state evolution information. Compared to traditional methods that rely on isolated interpretations based on single-section, single-moment results, this invention can continuously record the manifestation trajectory, scale changes, anomaly enhancement, and relationship transformation processes of the same structural face during face advancement, thus providing a direct basis for subsequent typical pattern recognition, anomaly expansion path extraction, and risk accumulation assessment.

[0200] Furthermore, because this step simultaneously retains the state information of newly added objects, objects not yet displayed, and terminated objects, the graph memory network is not a simple "overwrite update," but a dynamic memory structure with a clear historical inheritance logic and temporal evolution semantics. This structure is particularly suitable for analysis scenarios where tunnel surrounding rock structural surfaces exhibit intermittent display, local splitting, reconvergence, and risk accumulation in different excavation cycles, and can significantly improve the accuracy of cross-cycle structural continuity identification and engineering interpretability.

[0201] V. Cross-Cyclic Association Analysis and Typical Pattern Recognition

[0202] After completing cross-cycle node matching and graph memory network updates, the graph memory network retains the manifestation sequence, spatial trajectory, relationship changes, and state evolution information of the surrounding rock structure object in multiple excavation cycles. This step further performs cross-cycle association analysis based on the graph memory network to identify the extension patterns, propagation paths, local transformation relationships, and complex anomaly evolution patterns of the surrounding rock structure object during continuous construction. Unlike single-cycle analysis, which can only describe local anomalies at the current moment, this step focuses on mining the temporal structural semantics of "how different cycles continuously change, how different objects interact, and how local risks gradually form."

[0203] In one embodiment, cross-cycle association analysis includes at least the following: First, performing continuity analysis on the node sequence corresponding to the same memory identifier to identify the cross-cycle extension trajectory of the same structural surface; Second, tracking the relationship changes between different types of nodes in multiple consecutive cycles to identify local transformation chains between crack propagation, enhanced seepage, slab formation, and deformation anomalies; Third, performing path search on directed propagation edges and multi-type relationship edges in the graph memory network to identify abnormal expansion paths and risk transmission paths; Fourth, performing pattern induction on local subgraphs with recurring and similar evolutionary characteristics to form typical pattern recognition results.

[0204] 1. Cross-cycle continuity analysis

[0205] For the same structural surface, crack, seepage anomaly zone, or monitoring anomaly unit, if it has a unified historical inheritance identifier in multiple excavation cycles, it can be regarded as a continuous temporal manifestation of the same memory object. Preferably, the system first extracts the node sequence of the same memory object in different cycles and sorts it according to the cycle number to form an object-level temporal trajectory. By analyzing the changes in the position, scale, direction, and local neighborhood of the nodes in the temporal trajectory, it can be determined whether its manifestation process is characterized by continuous extension, local discontinuity, gradual enhancement, gradual attenuation, or bifurcation and convergence.

[0206] In a preferred embodiment, for any memory object Its manifest trajectory in multiple loops can be represented as: ;

[0207] in: , ,..., This indicates the excavation cycle number in which the object was made visible; This represents the corresponding cross-cycle manifestation trajectory. Based on this trajectory, further calculations can be made of indicators such as its continuous manifestation length, the number of manifestation interruptions, scale growth trend, and spatial advancement direction.

[0208] For structural surfaces such as joints, fissures, and bedding, if they have high directional consistency, spatial continuity, and scale continuity in adjacent or nearby cycles, they can be identified as "the same structural surface extending across cycles"; if they are briefly interrupted in some cycles, but the preceding and following cycles still maintain a strong continuity relationship, they can be identified as "the discontinuous manifestation of continuity"; if a historical object forms multiple sub-branches in subsequent cycles, it can be identified as "the structural bifurcation pattern"; if multiple historical fragments exhibit a unified structure in a new cycle, it can be identified as "the structural convergence pattern".

[0209] 2. Analysis of Local Transformation Relationships

[0210] In actual evolution, surrounding rock anomalies do not always remain unchanged in their object category. Instead, they often manifest as one type of anomaly gradually triggering or transforming into another. For example, local fractures gradually expand from their initial development to the instability of the block boundary, eventually forming a block collapse zone; similarly, early isolated seepage points continuously strengthen and expand along structural planes in multiple cycles, eventually forming a continuous water-rich anomaly zone. Therefore, this embodiment preferably analyzes the cross-cycle associations between nodes of different categories in a graph memory network to identify the local transformation relationships between anomaly objects.

[0211] In one embodiment, the system examines the order of appearance, distance variation trend, intensity variation trend, and relationship persistence of heterogeneous node pairs with directed risk propagation edges or multi-cycle co-occurrence relationships in multiple consecutive cycles. If a source node continuously strengthens in the first few cycles, and a target node subsequently appears and remains active in its neighborhood, it can be determined that there is a transformation or induction relationship from source anomaly to target anomaly.

[0212] Preferably, the local conversion intensity can be expressed as: ;

[0213] in, Indicates the first The strength of the propagation or coupling relationship between the source node and the target node in each loop. Indicates that the source node is at the . The degree of abnormal activity in each cycle, This indicates the degree to which the target node becomes more visible or enhanced in the next iteration. Represents a node To the node The intensity of cross-cycle conversion. When When the threshold is exceeded, a stable abnormal transformation relationship can be determined.

[0214] Based on the above mechanism, at least the following typical transformations can be identified:

[0215] (1) The crack transforms into a drop zone, that is, the crack node continues to grow in the preceding cycle, and then the drop node appears in its neighboring area and is accompanied by boundary expansion;

[0216] (2) The seepage points extend into the water-rich zone, that is, multiple discrete seepage nodes gradually connect along the direction of the same group of structural surfaces and form a continuous regional anomaly.

[0217] (3) Structural anomalies are coupled to deformation anomalies, that is, after the joint-dense area or crack propagation area is continuously active, the response intensity of the adjacent monitoring anomaly nodes increases significantly.

[0218] (4) The area of ​​rockfall becomes more unstable in the local area, that is, the area of ​​rockfall expands in subsequent cycles, and new cracks or abnormal deformation nodes are added around it.

[0219] 3. Anomaly Expansion Path Analysis

[0220] In graph memory networks, different nodes and relational edges can collectively form anomaly propagation paths that unfold along time and space. An anomaly propagation path refers to a continuous link from the initial manifestation and local enhancement of a certain type of surrounding rock anomaly to its propagation to neighboring areas. This path may manifest as a temporal extension of a single type of object, or as a successive transformation and joint diffusion between multiple types of objects.

[0221] In a preferred embodiment, the system constructs a candidate path set based on directed propagation edges, geometric continuation edges, and cross-cycle inheritance edges in the graph memory network, and performs reachability search, weight accumulation, and temporal consistency verification on each path. If a path maintains consistent direction, stable relationships, and consistently high node activity across multiple consecutive cycles, it can be identified as an abnormal expansion path.

[0222] Preferably, a path starts from the starting node To the terminal node Cross-loop exception path The path strength can be written as: ;

[0223] in, Represents the edges in the path Relationship weights This represents the temporal persistence coefficient or confidence coefficient of the corresponding edge. This indicates the overall expansion strength of the path. The higher the path strength, the more stable and significant the abnormal expansion trend expressed by the path.

[0224] Based on the above path analysis, results such as "the path of the fracture zone extending from the sidewall of the working face towards the arch", "the path of the seepage anomaly extending from the joint group towards the local water-rich zone", and "the path of the risk of local rockfall extending from the sidewall towards the middle of the working face" can be extracted. These types of paths not only reflect the location where the anomaly has occurred, but also reflect its possible direction of continued development, which directly supports the subsequent determination of key areas of concern.

[0225] 4. Typical Pattern Recognition

[0226] To transform the results of cross-cycle correlation analysis into interpretable and referable engineering identification results, this embodiment further identifies typical patterns in a graph memory network. Typical patterns refer to local subgraph structures that repeatedly appear in multiple excavation cycles, possessing relatively stable structural characteristics and evolutionary patterns. By summarizing complex time-series graph relationships into a finite number of typical patterns, the clarity of the results is improved, and it facilitates subsequent integration with risk assessment rules, early warning rules, and construction response strategies.

[0227] In a preferred embodiment, the system identifies several predefined or self-learned typical patterns by matching subgraph structures, clustering path features, or comparing rule templates. Preferably, these patterns include at least the following types:

[0228] (1) Cross-cycle extension mode of the same structural plane

[0229] This model corresponds to a situation where the same set of joints, fissures, or bedding objects continuously appear in similar directions across multiple consecutive cycles, and their location advancement is coordinated with the tunnel face's advance direction. This model typically indicates that a structural plane is not locally isolated but continuously controls the surrounding rock structure during excavation. For this model, the system can further output its extension direction, number of consecutive cycles, the trend of its length increase, and changes in local risk level.

[0230] (2) Transformation mode of cracks into spalling areas

[0231] This pattern corresponds to a situation where fracture nodes exhibit increased length, aperture, or density over multiple consecutive cycles, followed by the appearance of block breakage nodes at their ends or in their vicinity, forming a stable directional propagation link. This pattern indicates that the local surrounding rock has evolved from a structural plane development stage to a block instability stage, typically posing a high engineering risk.

[0232] (3) The pattern of seepage point extending to water-rich zone

[0233] This model corresponds to the gradual increase, strengthening, and connection of multiple discrete seepage nodes along a certain structural control direction over several cycles, eventually forming a banded distribution or regional seepage anomaly zone. This model reflects the gradual establishment and expansion of seepage channels, and is often associated with softening of surrounding rock, reduced local stability, and disturbance of subsequent support.

[0234] (4) Structural anomaly-water anomaly-deformation anomaly coupling mode

[0235] This model corresponds to a situation where joint and fissure development is enhanced, seepage activity increases, and monitoring anomalies are simultaneously amplified within a certain region. This model indicates that surrounding rock anomalies are no longer a single structural problem, but rather a complex risk pattern characterized by the coupled evolution of geological structure, groundwater action, and mechanical response. In engineering, this model should generally be considered a high-priority focus.

[0236] (5) Local abnormality recurrence pattern

[0237] This pattern corresponds to situations where anomalous nodes in a local area may not be continuously displayed in every cycle, but they recur in multiple interval cycles, and their neighborhood structures are highly similar. This pattern is suitable for identifying intermittent exposure anomalies caused by acquisition occlusion, local cleanup, or changes in display conditions, and helps avoid misjudging them as independent new anomalies.

[0238] 5. Pattern credibility and result selection

[0239] Because the data collected on-site may contain noise, missing data, or local disturbances, this embodiment preferably calculates the pattern credibility for each recognition result to improve the reliability of typical pattern recognition results. Pattern credibility can comprehensively consider the following factors: the number of nodes participating in the pattern, the stability of relation edges, the duration of the time series, the degree of support from multi-source data, the degree of path repetition, and the consistency of local structures. Patterns with high credibility are directly included in the formal analysis results; patterns with medium credibility can be marked as patterns awaiting review; patterns consisting only of a single low-quality node or short-term weak relations may not be output or may be downgraded for display.

[0240] Preferably, the credibility of a typical pattern can be expressed as: ;

[0241] in, Representation pattern The number of valid nodes included. This represents the average strength of the edges representing relationships within the pattern. Indicates the duration of the loop in the pattern. This indicates the support or overall quality evaluation value of multi-source data. These are weighting coefficients. This method allows for the sorting and filtering of results from different modes, enhancing the engineering practicality of the final output.

[0242] 6. The Implementation Effect of this Step

[0243] Through this step, the nodes, edges, and temporal states in the graph memory network no longer exist merely as raw records, but are further organized into continuous results, transformation relationship results, extended path results, and typical pattern results with clear engineering significance. Compared with traditional methods, this invention can not only answer "What anomalies are there at the current working face?", but also further answer "Are these anomalies related to previous cycles?", "Are they expanding or transforming?", and "Which local anomalies have formed stable risk patterns?"

[0244] In particular, for typical situations such as cross-cycle extension of the same structural surface, transformation of cracks into spalling areas, expansion of seepage points into water-rich zones, and enhanced coupling between structural anomalies and monitoring anomalies, this invention can perform continuous identification and structured representation on a graph memory network, thereby providing a direct basis for subsequent risk accumulation assessment, delineation of key areas of concern, and construction intervention decisions. This step is also an important manifestation that distinguishes this invention from single-cycle identification methods and static anomaly interpretation methods.

[0245] VI. Risk Accumulation Assessment and Determination of Key Areas of Concern

[0246] After completing cross-cycle correlation analysis and typical pattern recognition, the graph memory network already contains the continuous manifestation trajectory, local transformation relationships, abnormal expansion paths, and typical pattern recognition results of the surrounding rock structure object in multiple excavation cycles. This step further conducts a risk accumulation assessment based on this graph memory network, and, combined with the risk spatial distribution, expansion trend, and local structural coupling characteristics, determines the surrounding rock areas that require key attention in subsequent construction. Unlike static judgments based solely on single observation results of the current cycle, this invention emphasizes incorporating both the "current abnormal state" and the "historical evolution process" into the risk assessment, making the risk results dynamic, cumulative, and traceable.

[0247] In one embodiment, the risk accumulation assessment comprehensively considers at least the following factors: the strength of the node's own anomaly, the number of times the node continuously manifests in multiple cycles, the growth trend of node attributes, the number of high-risk relation edges in the node's neighborhood, the stability of the anomaly expansion path, the credibility of typical patterns, and the coupling degree between structural anomalies and monitored anomalies. Through the above multi-dimensional information fusion, the present invention can integrate originally scattered local anomalies into a risk state expression with continuous evolutionary semantics, thereby improving the accuracy of risk identification and engineering interpretability.

[0248] 1. Node-level risk accumulation assessment

[0249] In a preferred embodiment, risk accumulation is first calculated for each memory node in the graph memory network. Unlike the initial risk value of a node mentioned above, the node risk in this step no longer reflects only the local anomaly level of the object in the current cycle, but also reflects its frequency of appearance, duration, enhancement degree, and neighborhood propagation impact in historical cycles. For the same node, if it continuously appears in multiple cycles, its scale gradually increases, its anomaly intensity continuously increases, or it maintains stable coupling with high-risk objects, its accumulated risk should be significantly higher than that of a weak anomaly node that only appears briefly in a single cycle.

[0250] Preferably, node In the The cumulative risk value after one cycle can be expressed as: ;

[0251] in, This represents the current risk value of the node in the current loop. Indicates the intensity of the node's historical manifestation. Indicates the intensity of abnormal node evolution. This represents the contribution value of the path or pattern propagation of the node. These are the weighting coefficients.

[0252] Among them, historical intensity The intensity of anomalous evolution is used to characterize the sustained activity of a node in its preceding cycle, and is preferably determined by a combination of cumulative occurrence count, consecutive occurrence length, and recent occurrence density; Used to characterize the growth trend of node attributes across multiple cycles, such as increased crack length, increased aperture, increased seepage intensity, expanded spalling area, or increased monitoring anomaly rate; path or pattern propagation contribution value. This is used to characterize whether the node is located in a high-intensity anomalous propagation path or a high-confidence typical mode. If it is located in a stable propagation link or a composite anomalous coupling mode, its contribution value will be increased accordingly.

[0253] In one embodiment, if a crack node, although currently of moderate risk, has appeared in multiple consecutive cycles with a continuously increasing length and aperture, and a new spalling node has recently appeared in its adjacent area, then the cumulative risk of this crack node will be significantly higher than that of a crack node that appears only for the first time in this cycle and does not show an expanding trend. Similarly, if a seepage node continuously strengthens and extends along the structural surface in multiple cycles, then even if its current seepage area has not reached its maximum value, its cumulative risk can still be determined to be of a high level.

[0254] 2. Risk assessment of border-level transmission

[0255] In addition to the risk accumulation of the nodes themselves, the relational edges in the graph memory network also carry risk information. In particular, geometric continuation edges, locally co-occurring edges, and risk propagation edges not only describe the connections between objects but also reflect the ability of anomalies to propagate and spread within the local structure. Therefore, this embodiment preferably further assesses edge-level propagation risk to identify which pairs of objects have formed stable risk interaction links.

[0256] In a preferred embodiment, for any memory edge The propagation risk value can be calculated comprehensively based on edge weights, duration of cycles, source node risk level, and target node sensitivity. Preferably, the propagation risk value can be expressed as: ;

[0257] in, Indicates the edge at the th The relation weights after each cycle This indicates the length of the cycle in which the edge persists. This represents the cumulative risk value of the source node. This indicates the vulnerability or sensitivity of the target node. In this way, if a propagation edge exists for a long time, has a high weight, and connects a high-risk source node and a highly sensitive target node, the propagation risk value of that edge will be significantly increased.

[0258] For example, in the propagation chain of "crack → rockfall zone", if the crack node continues to grow in multiple cycles and the boundary of the rockfall zone node expands significantly, the propagation edge connecting the two will be given a high risk; in the propagation chain of "water seepage → abnormal deformation", if the increase in water seepage and the increase in displacement continue to be synchronized, the propagation risk of the corresponding directed edge will also increase significantly.

[0259] 3. Local submap and regional risk aggregation

[0260] In practical engineering, risks are often not determined by a single node or edge, but rather manifest as complex risks formed by the combined effects of multiple abnormal nodes and relational edges within a local area. Therefore, this embodiment further aggregates risks in the local subgraph or spatial neighborhood based on the node-level and edge-level risk results to form regional-level risk results.

[0261] In a preferred embodiment, the system uses a local window at the tunnel face, a regular grid, a spatial clustering unit, or a key structural partition as a regional analysis carrier. It fuses the node risk values ​​and edge propagation risk values ​​falling within the same region to obtain the comprehensive risk level of the corresponding region. Preferably, the region... The overall risk value can be expressed as: ;

[0262] in, This represents the risk contribution coefficient of nodes within the region. This represents the contribution coefficient of intra-regional transmission risk. Indicates the region In the The comprehensive risk value after each cycle. This aggregation method can map various information such as localized crack density, active seepage, spalling expansion, and enhanced monitoring anomalies onto a unified regional risk.

[0263] Preferably, for local areas where structural anomalies, water seepage anomalies, and deformation anomalies occur simultaneously, the coupling coefficient can be increased or a composite enhancement factor can be set when performing risk aggregation to highlight the higher risk of composite anomaly areas compared to single anomaly areas.

[0264] 4. Risk Level Classification and Dynamic Adjustment

[0265] After obtaining the risk values ​​at the node, edge, and region levels, the system classifies the risk levels for each object and region. Preferably, the risk levels can be divided into four levels: low risk, medium risk, high risk, and extremely high risk. Further refinement to more levels is also possible based on project requirements. The level thresholds can be determined using historical statistical data and empirical rules, or they can be calibrated by combining previously observed anomaly samples.

[0266] In one embodiment, the risk level is not fixed but can be dynamically adjusted as the excavation cycle progresses. For example, a local area may only exhibit medium risk in a single cycle, but if the risk continues to increase in multiple consecutive cycles, or its associated path lengthens significantly, its risk level can be upgraded to high risk. Conversely, if a local anomaly has occurred in a certain area but continues to weaken in subsequent cycles without forming a propagation link, its risk level can be downgraded. Through this dynamic adjustment mechanism, the risk assessment results can better reflect the actual state changes during the construction process.

[0267] Furthermore, the system can set differentiated risk assessment rules based on different object categories. For example, stricter high-risk triggering conditions can be used for areas with spalling and abnormal deformation units; for initial isolated cracks or weak seepage points, it is necessary to consider their historical evolution trends before deciding whether to upgrade the risk level. This avoids over-warning for short-term minor anomalies while promptly identifying objects with continuously accumulating hidden dangers.

[0268] 5. Focus on determining key areas

[0269] After completing the risk level classification, this step further identifies the surrounding rock areas that require key attention during subsequent construction. These key attention areas refer to spatial regions that exhibit high cumulative risk, a stable expansion trend, or significant coupling characteristics after the current cycle, and are likely to continue developing, expanding, or inducing new anomalies during subsequent excavation. This area is not limited to the anomalies already exposed at the current working face, but also includes potential influence zones along the structural continuity direction, propagation path direction, or the expansion direction of local high-risk sub-maps.

[0270] In a preferred embodiment, the determination of a key area of ​​concern is based on at least the following conditions: First, the overall risk value of the area is higher than a preset concern threshold; second, the area contains high-risk nodes or high-propagation-risk edges; third, the area is located on a stable abnormal expansion path or is an important component of a highly reliable typical pattern; fourth, the area repeatedly appears in multiple consecutive cycles or shows a significant increasing trend; and fifth, the area has a spatial proximity to the subsequent excavation direction, weak support areas, or key construction parts.

[0271] Preferably, the key concern area is not simply the union of the current high-risk grids, but rather spatially extrapolated based on the direction of anomaly expansion. For linear structural surfaces, a strip-shaped key concern area can be formed along their extension direction; for seepage and spalling expansion patterns, a strip-shaped or fan-shaped concern area can be formed along their propagation path; for localized complex anomaly areas, a buffer concern range can be formed around the area. In this way, the key concern area not only reflects "where the current risks are," but also "where the risks that may be prioritized in the next stage are."

[0272] 6. Focus on object sorting and output priority determination.

[0273] To facilitate field application, this embodiment preferably sorts the key concern objects based on the determination of key concern areas. The sorted objects may include high-risk nodes, high-risk edges, typical pattern subgraphs, and regional units. The sorting criteria may comprehensively consider factors such as cumulative risk value, propagation potential, pattern credibility, proximity to key construction locations, and the scope of subsequent impact.

[0274] For example, if a certain area has a slightly lower current risk value than another local area, but it is in a "structural anomaly-water anomaly-deformation anomaly coupled mode" and has a clear expansion path along the excavation front, then its output priority can be higher than that of an area that only shows a single rockfall anomaly but lacks signs of continuous expansion. Through this sorting mechanism, the system can provide a more targeted priority list for the construction site, facilitating subsequent monitoring intensification, support reinforcement, and risk management.

[0275] 7. The Implementation Effect of this Step

[0276] Through this step, the continuously manifested information, relational propagation information, and typical pattern information in the aforementioned graph memory network are further transformed into directly applicable risk level results and spatial concern results. Compared with the traditional method that relies on single-cycle static interpretation and manual experience thresholds, this invention can integrate node risk, edge propagation risk, regional coupling risk, and historical evolution trend into the evaluation process, thereby achieving dynamic cumulative determination of surrounding rock risk.

[0277] Especially for typical situations such as the continuous extension of the same structural surface, the transformation of cracks into spalling areas, the expansion of seepage points into water-rich zones, and the enhanced coupling between structural anomalies and monitoring anomalies, this invention can not only identify their existence but also quantify their risk contribution, assess their development level, and further delineate the local areas and expansion directions that should be given priority in subsequent construction. This allows the risk assessment results to go beyond a simple determination of "whether it is abnormal" and instead output structured conclusions such as "how great is the risk, how is it spreading, and where should the focus be next," thus providing direct basis for subsequent results output, construction early warning, and on-site decision-making.

[0278] VII. Result Output and Continuous Incremental Maintenance of Graph Memory

[0279] After completing cross-cycle correlation analysis, typical pattern recognition, and risk accumulation assessment, the system needs to output the analysis results in a queryable, traceable, visualized, and callable manner, while simultaneously performing incremental maintenance on the graph memory network. Unlike methods that only output the results of a single identification at the current moment, this invention emphasizes that the output should not only reflect the surrounding rock structure state under the current excavation cycle, but also its continuous relationship with previous cycles, abnormal expansion paths, risk change trends, and subsequent attention recommendations. Simultaneously, the system should stably write newly added nodes, relationships, states, and conclusions in the current cycle into the graph memory network, providing a foundation for continued cross-cycle analysis in subsequent cycles.

[0280] In one embodiment, this step includes at least the following processing: First, uniformly encapsulating node-level results, edge-level results, path-level results, pattern-level results, and region-level results; Second, generating corresponding output formats according to different uses such as face display, risk warning display, historical trajectory query, and result call interface; Third, fusing and storing the current cycle graph structure with historical graph memory, completing state updates, version records, and historical archiving; Fourth, when a new cycle arrives, continuing incremental analysis based on the maintained graph memory network without having to repeatedly construct all historical relationships.

[0281] 1. Organization and Structured Encapsulation of Results

[0282] In a preferred embodiment, the system first categorizes and organizes the results of the current loop and cross-loop analyses. The results include at least the following:

[0283] (1) The results of structural surface continuity analysis are used to characterize the extension, discontinuity, bifurcation or convergence of the same joint, crack or bedding object in multiple excavation cycles;

[0284] (2) Anomaly propagation path results are used to characterize the propagation links of cracks, seepage, spalling and deformation anomalies along time and space;

[0285] (3) Typical pattern recognition results are used to characterize patterns such as the transformation of cracks into spalling areas, the expansion of seepage points into water-rich zones, and the enhanced coupling between structural anomalies and monitoring anomalies.

[0286] (4) Risk accumulation level results, used to characterize the risk level of nodes, edges, local areas and key areas of concern;

[0287] (5) Key focus area results are used to indicate the local areas of the tunnel face and the potential impact zone ahead that should be given priority in subsequent construction.

[0288] To facilitate internal system processing and external invocation, this embodiment preferably encapsulates the above results into a unified result object. Each result object includes at least the following fields: result category, associated node or region identifier, cycle number, spatial location range, result value, reliability level, and generation time. This unified encapsulation ensures that different modules, terminals, and management platforms can all access data in the same format.

[0289] Preferably, for a certain output entity , can be represented as: ;

[0290] in, Indicates the result category, Indicates the associated object or region identifier. Indicates spatial location or spatial range. This indicates the corresponding analysis result value. Indicates the level of credibility. This indicates the excavation cycle number. This unified format enables structured storage and standardized interface output.

[0291] 2. Visualization of Results Output

[0292] In one embodiment, the system outputs the analysis results to a working face visualization interface, a construction management terminal, or an early warning platform. Preferably, the visualization output includes at least the following forms:

[0293] Firstly, the tunnel face structure continuity display diagram: This diagram uses the tunnel face image, point cloud unfolded diagram, or engineering coordinate plane as the base map, and marks continuously appearing joints, fissures, and bedding objects with a unified historical identifier or color, showing their extension direction, number of consecutive appearances, and state change trends. Through this diagram, on-site personnel can intuitively see which structural surfaces are not single-time exposure anomalies, but rather persist and progress through multiple cycles.

[0294] Secondly, the abnormal expansion path display diagram: This diagram shows the coupling links of crack expansion, water seepage propagation, rockfall evolution and monitoring anomalies in the form of lines, arrows or paths, which makes it easy to identify how local anomalies gradually develop from early signs to significant risks.

[0295] Third, the risk level distribution map: This display map uses color grading to express the risk level of each local area, key object and key path at the current working face, and can overlay high-risk nodes, high-propagation-risk edges and key concern area boundaries to achieve rapid identification of the spatial distribution of risk.

[0296] Fourth, the key focus area prompt map: This display map highlights the local areas that require priority attention, structural extension zones, or potential impact zones ahead, and can simultaneously mark the suggested monitoring densification zones, support reinforcement zones, or review and inspection zones.

[0297] In a preferred embodiment, the system also supports multi-level display of results, allowing switching between "object layer—relationship layer—path layer—region layer." The object layer is used to view individual joints, cracks, seepage points, spalling areas, or monitoring anomaly units; the relationship layer is used to view the adjacency, continuity, and propagation relationships between nodes; the path layer is used to view anomaly propagation paths and typical pattern links; and the region layer is used to view risk distribution and key areas of concern. This multi-level switching improves the flexibility and engineering adaptability of the results output.

[0298] 3. Query Results and Decision Support Output

[0299] In addition to graphical display, this embodiment preferably outputs structured query results and decision support results simultaneously. The structured query results can be indexed by node, cycle, region, pattern, and risk level, facilitating retrieval by on-site personnel, management personnel, or subsequent analysis modules. For example, information such as "the complete evolution record of a joint object from its first appearance to the current cycle," "the risk change trend of a certain region in the last five cycles," and "in which cycle did a certain high-risk pattern first form" can be queried.

[0300] The decision support outputs focus on converting analysis results into actionable engineering suggestions. For example, for high-risk fracture extension zones, the output might suggest "strengthening local verification and support attention"; for seepage extending into water-rich zones, the output might suggest "increasing the frequency of seepage monitoring and geological verification"; and for areas with coupled structural, water, and deformation anomalies, the output might suggest "designating these areas as key monitoring and early warning zones." The emphasis here is on providing areas of focus and priority, rather than replacing specific construction designs or on-site handling decisions.

[0301] 4. Write the current loop result and incrementally update the graph memory.

[0302] After the results are output, the system needs to write the newly formed nodes, relationship edges, state changes, path information, pattern results, and risk results in the current loop into the graph memory network to complete the continuous incremental maintenance of the graph memory. Unlike the traditional overwrite update method, the graph memory maintenance in this invention emphasizes "new additions can be written, history can be retained, state can be updated, and conclusions can be traced".

[0303] In a preferred embodiment, graph memory updating includes at least the following:

[0304] (1) Write newly appearing objects in the current loop that do not yet have a historical mapping into the memory node set and assign a new memory identifier;

[0305] (2) Write the established inheritance, splitting, merging, and termination relationships between the current loop and historical nodes into the mapping record;

[0306] (3) Write the newly formed or updated relation edges in the current loop into the memory edge set, and record their edge type, directionality, weight change and duration;

[0307] (4) Add the risk value, state value and confidence value of each node and edge in the current loop to its historical sequence;

[0308] (5) Write the typical patterns, expansion paths and key areas of focus identified in the current loop into the historical event database for subsequent querying and backtracking.

[0309] Preferably, the first The graph memory increment after each cycle can be represented as: ;

[0310] in, This represents a set of newly added or updated node information. This represents the set of newly added or updated edge information. This represents the set of state changes and mapping relationships. This represents the set of resulting events. In this way, the system does not rewrite the entire graph in each loop, but only writes the incremental changes relative to the previous loop, thereby reducing storage overhead and improving update efficiency.

[0311] 5. Historical version retention and traceable maintenance

[0312] To ensure the traceability and verifiability of the analysis process, this embodiment preferably retains historical version information during the continuous incremental maintenance of the graph memory. That is, the system not only saves the "latest state" but also the "stage state at the end of each cycle." This way, when it is necessary to review a particular abnormal evolution process later, the graph memory state under any historical cycle can be traced back to view the node display, relationship structure, risk level, and areas of interest at that time.

[0313] In one embodiment, after each cycle update, the system generates a graph state snapshot corresponding to the cycle number, while simultaneously retaining a critical event log. The critical event log includes at least the following events: the first appearance of a node, the disappearance of a node, anomaly level transitions, the first establishment of a risk propagation edge, the first formation of a typical pattern, and the first triggering of a key area of ​​concern. This maintenance method, combining snapshots and events, significantly enhances the system's ability to interpret the evolution of anomalies.

[0314] Preferably, to prevent the graph memory from expanding indefinitely with the number of cycles, the system can also be configured with a history compression and hierarchical storage mechanism. For nodes and edges that have not been displayed for a long time and have low relevance to the current analysis, only their key summary information can be retained, while the complete details are transferred to the history archive area; for recently active objects and high-risk patterns, the complete high-precision state is retained. Through this mechanism, long-term operating efficiency can be improved while ensuring historical traceability.

[0315] 6. Continuous incremental operation for subsequent new cycles

[0316] A key feature of this invention is that the graph memory update after the current loop completes is not the end point, but rather the starting point for the next loop analysis. When the... When new data arrives in a new excavation cycle, the system directly uses... Based on historical data, node matching, relationship updating, pattern recognition, and risk assessment can continue without reprocessing all historical data. In other words, the method of this invention is naturally applicable to continuous incremental operation in continuous construction scenarios.

[0317] In a preferred embodiment, upon receiving new cycle data, the system prioritizes active nodes, undisclosed nodes, and high-risk edges from the previous cycle as key matching targets, and performs incremental updates based on the latest data. For long-term stable, low-risk historical objects, their summary information is only retrieved when needed. Through this continuous incremental mechanism oriented towards new cycles, the system can improve real-time response capabilities while ensuring historical continuity.

[0318] 7. The Implementation Effect of this Step

[0319] Through this step, the present invention truly transforms the analysis results generated in the preceding steps into visible, verifiable, storable, and sustainably operational engineering application results. On the one hand, the system can output structural surface continuity analysis results, anomaly propagation paths, typical patterns, risk levels, and key areas of concern, enabling on-site personnel to clearly understand "what the current anomaly is, where it comes from, where it is going, and where it deserves the most attention." On the other hand, through incremental writing, historical retention, status updates, and version management, the system enables the graph memory network to continuously evolve with the excavation cycle without losing historical information due to the arrival of new data.

[0320] Compared to traditional methods that statically archive only single-cycle identification results, this invention enables continuous memorization of surrounding rock structure information, continuous tracking of anomaly evolution processes, and long-term accumulation of analysis results, thus making cross-cycle correlation analysis truly applicable in engineering. This result output and continuous incremental maintenance mechanism are also important foundations for this invention to adapt to long-distance tunnel continuous construction, complex surrounding rock dynamic changes, and multi-round risk evolution analysis.

[0321] Example 2:

[0322] In this embodiment, corresponding to the above-mentioned graph memory-based cross-cycle correlation analysis method for tunnel surrounding rock structural surfaces, a graph memory-based cross-cycle correlation analysis system for tunnel surrounding rock structural surfaces is constructed. This system is used to uniformly access, standardize, construct graph nodes, model relationships, update cross-cycle memories, perform correlation analysis, conduct risk assessment, and output results for the surrounding rock structural information of the tunnel face under multiple consecutive excavation cycles during tunnel construction. This enables continuous cross-cycle tracking of the surrounding rock structural surfaces, identification of abnormal evolution, and determination of key risk areas.

[0323] The system can be deployed on edge computing terminals, project servers, central servers, or cloud platforms. In scenarios with high real-time requirements at construction sites, it is preferable to deploy functions such as data access, standardized processing, preliminary node construction, and local risk identification on edge computing terminals. In scenarios requiring multi-cycle historical analysis, unified management of the entire line segment, or complex graph memory reasoning, it is preferable to deploy graph memory updates, cross-cycle association analysis, risk assessment, and comprehensive result output on project servers or cloud platforms. Data interaction between system modules can be achieved through bus mechanisms, service call mechanisms, message queue mechanisms, or database sharing mechanisms.

[0324] In one embodiment, the system includes at least: a data access module, a data alignment and standardization module, a node construction module, a relationship modeling module, a graph memory update module, a cross-cycle association analysis module, a risk assessment module, and a result output module. The modules organize data flow according to the excavation cycle number and collaborate around a unified graph memory database. Each module is described in detail below.

[0325] 1. Data Access Module

[0326] The data access module is used to access multi-source surrounding rock observation data related to the current excavation cycle and to collect the raw data according to a unified cycle number. The multi-source observation data includes at least the face image data, 3D point cloud data, geological sketches or geological logging data, seepage inspection records, rockfall records, and convergence, settlement, and displacement monitoring data uploaded by the monitoring system.

[0327] In one embodiment, the data access module may include an image access unit, a point cloud access unit, a text recording access unit, a monitoring data access unit, and a field event access unit. The image access unit receives images of the working face acquired by industrial cameras, handheld terminals, inspection terminals, or mobile robots; the point cloud access unit receives point cloud data generated by laser scanning equipment, structured light equipment, or photogrammetry devices; the text recording access unit receives information such as surrounding rock type, joint group, bedding information, seepage description, and rockfall description entered by geologists; the monitoring data access unit receives real-time or near-real-time observations from monitoring points; and the field event access unit receives supplementary event information related to the working face, such as local spalling, signs of sudden water inrush, and abnormal support.

[0328] Preferably, after receiving data, the data access module first assigns an excavation cycle identifier, a collection time identifier, an equipment source identifier, and a data type identifier to each data piece, and then performs a basic integrity check. For data that failed to be collected, is corrupted, has an abnormal format, or has an unknown source, it can be directly written to the exception log or enter the pending review area; for data that was successfully accessed, it is written to the raw data cache area to provide input for subsequent standardized processing.

[0329] 2. Data Alignment and Standardization Module

[0330] The data alignment and standardization module is connected to the data access module and is used to perform unified time alignment, spatial coordinate unification, scale normalization, field structuring, and quality assessment on the accessed multi-source raw data to generate standardized input data suitable for graph modeling.

[0331] In one embodiment, the data alignment and standardization module may include a time synchronization unit, a spatial registration unit, an attribute normalization unit, a text structuring unit, and a quality control unit. The time synchronization unit is used to unify the monitoring curve, image sampling time, point cloud sampling time, and manual recording time with reference to the end time of the current excavation cycle or a specified reference time. The spatial registration unit is used to map image coordinates, point cloud coordinates, measuring point coordinates, and engineering coordinates to a unified local coordinate system of the tunnel face or a tunnel engineering coordinate system. The attribute normalization unit is used to unify the scale of data with different dimensions such as length, area, displacement, and seepage intensity. The text structuring unit is used to extract fields such as surrounding rock category, structural surface group, anomaly category, and severity from geological logging, inspection records, and manual notes. The quality control unit is used to identify blurred images, missing point clouds, abnormal jump monitoring values, and incomplete text records, and generate corresponding quality labels and confidence level indicators.

[0332] Preferably, after processing, this module organizes the multi-source data of the current excavation cycle into standardized data entries. Each data entry includes at least the cycle number, object category, spatiotemporal location, geometric parameters, attribute parameters, data source, and quality evaluation value. In this way, the differences between different data sources are uniformly processed before entering the node construction.

[0333] 3. Node building module

[0334] The node construction module is used to identify surrounding rock structural objects based on standardized input data and convert joints, fissures, bedding, seepage points, spalling zones, and deformation anomaly units into graph-structured nodes. Graph nodes serve as the basic carrier for subsequent relationship modeling, cross-cycle matching, and risk propagation analysis.

[0335] In one embodiment, the node construction module may include a structural object extraction unit, a regional anomaly identification unit, a monitoring anomaly mapping unit, and a node encoding unit. The structural object extraction unit is used to extract linear or planar structural objects such as joints, cracks, and bedding; the regional anomaly identification unit is used to extract regional anomaly objects such as seepage points, spalling areas, and locally fractured areas; the monitoring anomaly mapping unit is used to map monitoring results such as convergence anomalies, settlement anomalies, and displacement anomalies to corresponding anomaly nodes; and the node encoding unit is used to uniformly encode the above objects into a node data structure.

[0336] Preferably, each node includes at least a node identifier, object category, cycle number, spatial location, geometric features, attribute features, quality evaluation value, and initial risk value. For joint, fracture, and bedding nodes, the direction, length, aperture, and group attributes can be recorded in detail; for seepage points and spalling areas, the center location, area range, area, and severity can be recorded in detail; for deformation anomaly nodes, the measuring point number, anomaly amplitude, and rate of change can be recorded in detail. For objects that need to enter cross-cycle continuous tracking, the node encoding unit can also reserve historical mapping fields and inherited state fields for them.

[0337] In one embodiment, after the node building module completes its processing, it generates a set of nodes for the current excavation cycle and writes the set of nodes into the current cycle graph cache to provide input for the relationship modeling module.

[0338] 4. Relationship Modeling Module

[0339] The relationship modeling module is used to construct various types of relationship edges between nodes based on the node set of the current excavation cycle, forming a single-cycle structure graph. Relationship edges include at least spatial adjacency, same-set joint, geometric continuation, local co-occurrence, and risk propagation relationships.

[0340] In one embodiment, the relationship modeling module may include an adjacency analysis unit, a phylogenetic discrimination unit, a continuation relationship analysis unit, a co-occurrence relationship analysis unit, and a propagation relationship modeling unit. The adjacency analysis unit generates spatial adjacency edges based on the spatial distance between nodes, object scale, and local distribution density; the phylogenetic discrimination unit generates joint edges within the same phylogenetic group based on the main direction, attitude similarity, and phylogenetic records; the continuation relationship analysis unit generates geometric continuation edges based on directional consistency, boundary continuity, and projective connectivity; the co-occurrence relationship analysis unit generates local co-occurrence edges based on the simultaneous occurrence of multiple anomalies within a local area; and the propagation relationship modeling unit generates directed risk propagation edges based on preset engineering mechanism rules and node state relationships.

[0341] Preferably, each relation edge includes at least the starting node identifier, the target node identifier, the edge type, the edge weight, the directionality marker, and the confidence evaluation value. Multiple relation edges are allowed to exist simultaneously for the same pair of nodes. The module outputs a single-cycle structure diagram corresponding to the current excavation cycle, which can be directly written to the current graph storage area for use by the graph memory update module.

[0342] 5. Graph memory update module

[0343] The graph memory update module is used to connect the single-cycle structure graph formed by the current excavation cycle with the historical graph memory network, establishing cross-cycle node inheritance relationships, relation edge update relationships, and state evolution relationships to form a new graph memory network. This module is one of the key modules for achieving "continuous cross-cycle expression".

[0344] In one embodiment, the graph memory update module may include a candidate filtering unit, a cross-cycle matching unit, a state update unit, a relation inheritance unit, and a historical version maintenance unit. The candidate filtering unit filters candidate objects that may correspond to the current node from historical memory nodes; the cross-cycle matching unit performs similarity matching between the current node and historical nodes by integrating spatial location, geometric features, attribute changes, and neighborhood structure; the state update unit inherits the historical identifier of successfully matched nodes and updates their state vector, appearance count, duration, and risk evolution fields; the relation inheritance unit aligns, incrementally updates, or deactivates relation edges in the current cycle with historical edges; and the historical version maintenance unit saves a snapshot of the graph state after each cycle update and a log of key events.

[0345] Preferably, the graph memory update module handles not only one-to-one continuation relationships, but also one-to-many splitting and many-to-one convergence relationships. For nodes newly appearing in the current loop that cannot be matched with historical objects, the module creates a new memory identifier for them; for objects in the history that are not yet displayed in the current loop, their historical records are retained and updated to the "not yet displayed" state, rather than being directly deleted. In this way, the graph memory update module can maintain the continuous historical semantics of surrounding rock structural surfaces, anomaly zones, and monitored anomaly objects.

[0346] 6. Cross-Cyclic Association Analysis Module

[0347] The cross-cycle correlation analysis module is used to analyze the continuous manifestation, local transformation, anomaly propagation, and complex coupling processes of surrounding rock structures in multiple excavation cycles based on the updated graph memory network, and to identify typical evolution patterns. This module corresponds to S5 in the method embodiment.

[0348] In one embodiment, the cross-cycle association analysis module may include a continuity analysis unit, a transformation relationship identification unit, a path analysis unit, and a pattern recognition unit. The continuity analysis unit is used to extract the manifestation sequence of the same memory object in multiple cycles and determine whether it is a continuous extension, discontinuous manifestation, bifurcation manifestation, or convergence manifestation. The transformation relationship identification unit is used to analyze the cross-cycle transformation relationships between dissimilar objects such as the transformation from cracks to spalling areas, the expansion of seepage points to water-rich zones, and the coupling of structural anomalies to deformation anomalies. The path analysis unit is used to extract anomaly propagation paths, relationship continuation paths, and composite anomaly extension links. The pattern recognition unit is used to identify typical patterns based on preset templates, subgraph matching, or rule-based clustering methods.

[0349] Preferably, the module outputs at least the following analysis results: cross-cycle extension pattern of the same structural plane, transformation pattern of cracks into spalling areas, expansion pattern of seepage points into water-rich zones, coupling pattern of structural anomalies, water anomalies, and deformation anomalies, and recurring pattern of local anomalies. These results can be directly provided to the risk assessment module.

[0350] 7. Risk Assessment Module

[0351] The risk assessment module is used to perform cumulative risk assessment on the surrounding rock structure and its local areas based on the node states, relation edge states, abnormal expansion paths, and typical pattern results in the graph memory network, and to determine the key areas of concern in subsequent construction. This module corresponds to S6 in the method embodiment.

[0352] In one embodiment, the risk assessment module may include a node risk assessment unit, a propagation risk assessment unit, a regional risk aggregation unit, a classification unit, and a key focus determination unit. The node risk assessment unit calculates the cumulative risk value of each memory node, taking into account current risk, historical manifestation intensity, attribute growth trend, and path contribution. The propagation risk assessment unit calculates the risk intensity of relationship edges, particularly propagation edges. The regional risk aggregation unit merges node and edge risks within a local spatial range to form a regional risk value. The classification unit maps risk values ​​to low, medium, high, or extremely high risk levels. The key focus determination unit combines regional risk values, pattern credibility, and expansion direction to determine local areas and potential impact zones requiring focused attention during subsequent construction.

[0353] Preferably, the risk assessment module can also output a ranking of key concerns, which identifies the structural surfaces, abnormal paths, and local areas that require the highest priority for review, monitoring, or early warning after the current cycle ends. This ranking can be determined comprehensively based on risk value, propagation potential, persistence, and proximity to critical construction sites.

[0354] 8. Result Output Module

[0355] The results output module is used to uniformly encapsulate, visualize, output structured queries, and continuously store the analysis results from the aforementioned modules. This module corresponds to S7 in the method embodiment and is an important module for realizing engineering applications.

[0356] In one embodiment, the result output module may include a graphical display unit, a structured result output unit, an interface service unit, and a historical archiving unit. The graphical display unit overlays structural continuity results, anomaly propagation paths, risk level distributions, and key areas of concern onto the working face image, point cloud unfolded diagram, or engineering coordinate plane diagram. The structured result output unit generates standardized result files indexed by node, region, path, pattern, and cycle. The interface service unit provides data access interfaces to construction management platforms, early warning systems, digital twin systems, or mobile terminals. The historical archiving unit saves graphical memory snapshots, risk results, pattern results, and key event logs for each cycle.

[0357] Preferably, the results output module supports multi-level display modes, including object layer, relationship layer, path layer, and region layer. The object layer is used to view individual joints, cracks, seepage points, spalling areas, or monitor abnormal nodes; the relationship layer is used to view various relationship edges; the path layer is used to view cross-loop extension paths and transformation links; and the region layer is used to view risk distribution and key areas of concern. Through layered display, on-site personnel can quickly switch analysis perspectives according to their needs.

[0358] Furthermore, the results output module can also write the newly added nodes, update relationships, state changes, risk levels, and typical patterns of the current cycle into the graph memory database, achieving synchronous updates between the system analysis results and the historical memory state. In this way, when the next excavation cycle data arrives, the system can directly continue incremental analysis based on the existing history without having to rebuild the entire historical graph structure.

[0359] System collaborative working process:

[0360] In a complete workflow, the data access module first receives data from the field acquisition terminal, scanning equipment, monitoring equipment, and manual data entry terminal. Multi-source surrounding rock data for each excavation cycle; the data alignment and standardization module performs temporal and spatial unification, attribute normalization, and quality assessment on the above data; the node construction module converts the standardized data into surrounding rock structure object nodes; the relationship modeling module constructs a single-cycle structure diagram based on the spatial, group, continuity, and propagation relationships between nodes; the graph memory update module matches and merges this single-cycle structure diagram with the historical graph memory network to form a new memory state; the cross-cycle correlation analysis module identifies patterns such as cross-cycle extension of the same structural surface, transformation of fractures into rockfall areas, and expansion of seepage into water-rich zones; the risk assessment module calculates node-level, edge-level, and regional-level risks and delineates key areas of concern; finally, the results output module outputs the analysis results to the visualization interface, management platform, and historical storage area, while simultaneously completing the incremental writing of the current cycle graph memory.

[0361] Through the modular system architecture described above, this embodiment enables continuous analysis and dynamic management of the cross-cycle evolution process of the tunnel surrounding rock structure. Compared with traditional systems that only perform static identification of a single-cycle tunnel face, this embodiment can retain historical structural memory, express anomaly propagation paths, quantify risk accumulation processes, and form key focus area results that are connected with subsequent construction processes at the system level. Therefore, it is more suitable for continuous monitoring, dynamic early warning, and intelligent auxiliary analysis of tunnel construction sites under complex geological conditions.

[0362] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for cross-cycle correlation analysis of tunnel surrounding rock structural surfaces based on graph memory, characterized in that, Includes the following steps: S1. Acquisition and unified representation of multi-source surrounding rock observation data: Acquire multi-source surrounding rock observation data under multiple consecutive excavation cycles of the tunnel. The multi-source surrounding rock observation data includes at least tunnel face image data, three-dimensional point cloud data, geological logging data, seepage observation data, rockfall area record data, and deformation monitoring data. The multi-source surrounding rock observation data is uniformly numbered, time-aligned, spatially registered, attribute-normalized, and structured-coded to form standardized input data. S2. Extraction of surrounding rock structure objects and construction of graph nodes: Based on the standardized input data, extract joints, fissures, bedding, seepage points, spalling areas and abnormal deformation units, and encode the extraction results into graph nodes. Each graph node includes at least the object category, the excavation cycle to which it belongs, the spatial location, the geometric features, the attribute features, the quality evaluation value and the initial risk value. S3. Relationship edge construction and single-cycle structure graph generation: Based on the spatial adjacency relationship, same group joint relationship, geometric continuity relationship, local co-occurrence relationship and risk propagation relationship between graph nodes, relationship edges are constructed to generate the single-cycle structure graph corresponding to the current excavation cycle; S4. Cross-cycle node matching and graph memory network update: Perform cross-cycle node matching between the single-cycle structure graph corresponding to the current excavation cycle and the graph memory network corresponding to the historical excavation cycles, establish inheritance, splitting or merging relationships between the current node and the historical node, and update the node state and relationship edge state to form the updated graph memory network. S5. Cross-cycle correlation analysis and typical pattern recognition: Based on the updated graph memory network, cross-cycle correlation analysis is performed to identify typical patterns such as cross-cycle extension of the same structural surface, transformation of cracks into spalling areas, expansion of seepage points into water-rich zones, and enhanced coupling between structural anomalies and monitoring anomalies, and to extract the anomaly expansion path. S6. Accumulated Risk Assessment and Determination of Key Areas of Concern: Based on the current risk of nodes, historical manifestation intensity, abnormal evolution intensity, and path or pattern propagation contribution, the cumulative risk of nodes, relationship edge propagation risk, and regional risk are assessed, and key areas of concern are determined. S7. Result Output and Graph Memory Continuous Incremental Maintenance: Output the structural surface continuity analysis results, abnormal expansion paths, risk accumulation levels, and key areas of concern. Write the new nodes, update relationships, state changes, risk results, and typical pattern results of the current excavation cycle into the graph memory network for continuous use in subsequent excavation cycles.

2. The method for cross-cycle correlation analysis of tunnel surrounding rock structural surfaces based on graph memory as described in claim 1, characterized in that: In step S1, multi-source surrounding rock observation data are collected using the excavation cycle number as the basic organizational unit; the time alignment includes interpolating and synchronizing continuous monitoring data using the end time of the current excavation cycle or the end time of the face clearing as the reference time; the spatial registration includes mapping image coordinates, point cloud coordinates, equipment local coordinates, and monitoring point coordinates to a unified local engineering coordinate system of the face; the structured coding includes recording the excavation cycle number, object category, acquisition time, spatial location, geometric features, anomaly intensity, data source, confidence level, and quality identifier for each observation object.

3. The method for cross-cycle correlation analysis of tunnel surrounding rock structural surfaces based on graph memory as described in claim 1, characterized in that: In step S2, the surrounding rock structure objects are divided into three categories: linear structure objects, regional anomaly objects, and monitoring response objects. The linear structure objects include joints, fissures, and bedding; the regional anomaly objects include seepage points and spalling areas; and the monitoring response objects include deformation anomaly units. The directional and geometric features of the linear structure objects, the center position and regional range features of the regional anomaly objects, and the anomaly amplitude and rate of change features of the monitoring response objects are extracted to generate the graph nodes.

4. The method for cross-cycle correlation analysis of tunnel surrounding rock structural surfaces based on graph memory as described in claim 1, characterized in that: In step S3, the construction of the relation edges includes: Perform candidate adjacency filtering on the graph nodes in the current excavation cycle; Spatial adjacency edges are established based on the spatial distance between nodes; Establish joint edges in the same group based on directional and attribute similarity between linear structure objects; Geometric continuation edges are established based on directional consistency, endpoint distance, projection connectivity, and intermediate blocking conditions; Establish local co-occurrence edges based on the synchronous appearance of different types of abnormal objects within the same local analysis unit; A directed risk propagation edge is established based on the preset object category propagation rules, the anomaly strength of the source node, and the degree of influence of the neighborhood; The set of graph nodes and the set of relation edges are then organized into the single-cycle structure graph.

5. The method for cross-cycle correlation analysis of tunnel surrounding rock structural surfaces based on graph memory according to claim 1, characterized in that: In step S4, the cross-loop node matching includes: Based on object category consistency, spatial proximity, directional compatibility, and structural semantic rationality, candidate historical nodes are selected. The matching similarity between the current node and candidate historical nodes is calculated by combining location similarity, geometric feature similarity, attribute feature similarity, and neighborhood relationship similarity. When the matching similarity is greater than the preset matching threshold and reaches the maximum value among the candidate objects, it is determined that the current node and the corresponding historical node constitute an inheritance matching relationship; If one historical node corresponds to multiple current nodes, it is determined to be displayed as a split; if multiple historical nodes correspond to one current node, it is determined to be displayed as a merge. The historical node status is marked as one of the following: active, not yet displayed, or terminated.

6. The method for cross-cycle correlation analysis of tunnel surrounding rock structural surfaces based on graph memory according to claim 1, characterized in that: In step S5, the cross-cyclic association analysis includes: Continuity analysis is performed on the node sequences corresponding to the same memory identifier to identify cross-cycle extension trajectories of the same structural plane; The relationship changes of different types of nodes in multiple consecutive excavation cycles are tracked to identify local transformation chains between crack propagation, enhanced water seepage, spalling formation and deformation anomalies; Path search is performed on directed propagation edges and multi-type relation edges in graph memory networks to identify anomalous expansion paths and risk propagation paths; The subgraph structure is matched, the path features are clustered, or the rule templates are compared to form typical pattern recognition results; Among them, the typical modes include at least the cross-cycle extension mode of the same structural plane, the transformation mode of cracks into slab areas, the expansion mode of seepage points into water-rich zones, the coupling mode of structural anomalies, water anomalies, and deformation anomalies, and the mode of repeated manifestation of local anomalies.

7. The method for cross-cycle correlation analysis of tunnel surrounding rock structural surfaces based on graph memory according to claim 1, characterized in that: In steps S6 and S7, the determination of the key concern area includes: aggregating the node risk value and the relationship edge propagation risk value within the local window of the tunnel face, regular grid, spatial clustering unit, or key structural partition to obtain the comprehensive risk value of the area; determining the key concern area based on the comprehensive risk value of the area, the number of high-risk nodes within the area, the number of high propagation risk edges, the abnormal expansion path, the typical mode, and the enhancement trend of multiple consecutive excavation cycles; and spatially extrapolating the key concern area along the extension direction of the linear structural object, the direction of the seepage propagation path, or the direction of the block fall expansion. The output results and continuous incremental maintenance of the graph memory include: outputting the structural surface continuity analysis results, abnormal expansion paths, risk accumulation levels and key areas of concern, and writing newly formed nodes, relation edges, state changes, risk levels and typical patterns in the current excavation cycle into the graph memory network.

8. A system for implementing the cross-cycle correlation analysis of tunnel surrounding rock structural surfaces based on graph memory as described in any one of claims 1-7, characterized in that, include: The data access module is used to access face image data, 3D point cloud data, geological logging data, seepage observation data, rockfall area record data, and deformation monitoring data from multiple consecutive excavation cycles. The data alignment and standardization module, connected to the data access module, is used to perform unified numbering, time alignment, spatial registration, attribute normalization, and structured coding on the multi-source surrounding rock observation data. The node construction module, connected to the data alignment and standardization module, is used to extract joints, cracks, bedding, seepage points, spalling areas, and abnormal deformation elements, and encode them as graph nodes. The relationship modeling module, connected to the node construction module, is used to construct relationship edges and generate a single-loop structure graph based on spatial adjacency, same-group joint relationship, geometric continuity relationship, local co-occurrence relationship and risk propagation relationship. The graph memory update module, connected to the relationship modeling module, is used to match and merge the single-cycle structure graph corresponding to the current excavation cycle with the historical graph memory network, update the node state and relationship edge state, and form an updated graph memory network. The cross-cycle correlation analysis module, connected to the graph memory update module, is used to identify typical patterns such as cross-cycle extension of the same structural surface, transformation of cracks into spalling areas, expansion of seepage points into water-rich zones, and enhanced coupling between structural anomalies and monitoring anomalies, and to extract the anomaly expansion path. The risk assessment module, connected to the cross-cycle correlation analysis module, is used to cumulatively assess node risk, relationship edge propagation risk, and regional risk, and to determine key areas of concern. The results output module, connected to the risk assessment module, is used to output the structural continuity analysis results, abnormal expansion paths, risk accumulation levels, and key areas of concern. It also writes the incremental data of newly added nodes, update relationships, state changes, risk results, and typical pattern results corresponding to the current excavation cycle into the graph memory network.

9. The tunnel surrounding rock structural surface cross-cycle correlation analysis system based on graph memory according to claim 8, characterized in that: The graph memory update module includes: a candidate filtering unit, a cross-loop matching unit, a state update unit, a relationship inheritance unit, and a historical version maintenance unit; The candidate filtering unit is used to filter candidate objects corresponding to the current node from historical memory nodes. The cross-cycle matching unit is used to perform similarity matching between the current node and historical nodes by comprehensively considering spatial location, geometric features, attribute changes and neighborhood structure. The state update unit is used to inherit the historical identifier of successfully matched nodes and update the state vector, number of appearances, duration and risk evolution fields. The relation inheritance unit is used to align, incrementally update or deactivate relation edges in the current cycle with historical edges. The historical version maintenance unit is used to save the graph state snapshot and key event log after each excavation cycle update.

10. The tunnel surrounding rock structural surface cross-cycle correlation analysis system based on graph memory according to claim 8, characterized in that: The result output module includes: a graphical display unit, a structured result output unit, an interface service unit, and a historical archive unit; The graphical display unit is used to display the analysis results in a multi-level display mode of object layer, relationship layer, path layer and region layer. The structured result output unit is used to output standardized results by node, region, path, pattern and excavation cycle index. The interface service unit is used to provide result call interface to construction management platform, early warning system, digital twin system or mobile terminal. The historical archiving unit is used to save graph memory network snapshots, risk results, typical pattern results and key event logs, and supports subsequent excavation cycles to continue to perform incremental analysis on the basis of existing history.

Images