Deep foundation pit deformation monitoring method and system based on close-range photogrammetry

Abstract

The present application relates to the technical field of foundation pit engineering monitoring, and discloses a deep and large foundation pit deformation monitoring method and system based on close-range photogrammetry, wherein the deep and large foundation pit deformation monitoring method based on close-range photogrammetry comprises the following steps: receiving risk area data to generate a standardized image set; combining domain knowledge to perform adaptive three-dimensional reconstruction to output a three-dimensional grid model with semantic annotation; extracting a model deformation mode and performing semantic annotation to form a structured set; constructing a bidirectional information channel between a geometric domain and a knowledge domain based on this to create a knowledge-geometry collaborative measurement graph; using a multi-scale space-time memory structure to realize self-evolution learning and resource optimization allocation, outputting a resource allocation strategy and updating a monitoring knowledge base; performing risk assessment according to the deformation mode, semantic interpretation and updated knowledge base to produce intervention suggestions, and outputting hierarchical early warning information and multi-level decision support schemes; and the present application realizes automatic processing and optimization from data to decision.

Description

Technical Field

[0001] This invention relates to the field of foundation pit engineering monitoring technology, and more specifically, to a method and system for monitoring the deformation of deep and large foundation pits based on close-range photogrammetry. Background Technology

[0002] With the continuous advancement of urbanization, large-scale underground engineering projects and ultra-deep foundation pit projects are increasing, making the safety and accuracy of foundation pit deformation monitoring particularly important. Traditional foundation pit monitoring methods, such as measuring targets, inclinometers, and displacement gauges, can provide monitoring data at discrete points, but they are difficult to obtain global deformation information on the foundation pit surface, and manual operation is arduous and inefficient. In recent years, although new monitoring technologies such as laser scanning and UAV photogrammetry have been introduced, these methods often face problems such as insufficient accuracy, poor real-time performance, and high cost in complex environments. Especially in densely built-up areas or large foundation pits with depths exceeding 30 meters, there is still a lack of monitoring solutions that can simultaneously achieve high accuracy, high efficiency, and low cost.

[0003] More seriously, current foundation pit deformation monitoring systems generally lack the ability to effectively integrate geometric observation data with engineering knowledge. This results in a large volume of monitoring data but limited intelligence, hindering early identification and warning of potential risks. Specifically, this manifests as: a semantic gap between monitoring data and deformation mechanisms, preventing the system from automatically explaining the engineering causes behind deformation phenomena; a lack of intelligent adaptive capabilities in monitoring resource allocation, failing to dynamically adjust sampling density based on risk levels; and difficulty in accumulating and passing on monitoring experience, hindering the system's cognitive abilities from self-improving with increasing project experience. This disconnect between geometric data and engineering knowledge, coupled with the lack of self-evolutionary capabilities in monitoring systems, has become a key bottleneck restricting the improvement of foundation pit safety monitoring effectiveness. Summary of the Invention

[0004] This invention provides a method and system for monitoring the deformation of deep foundation pits based on close-range photogrammetry, which solves the technical problem of the separation between geometric data and engineering knowledge in related technologies.

[0005] This invention provides a method for monitoring deformation of deep foundation pits based on close-range photogrammetry, comprising the following steps:

[0006] Receive risk area data from the knowledge graph, perform adaptive acquisition and preprocessing operations, and generate a standardized image set;

[0007] By combining standardized image sets with domain knowledge constraints, adaptive 3D reconstruction guided by knowledge geometric interaction entropy is performed to generate a 3D mesh model with semantic annotations.

[0008] The deformation patterns of a 3D mesh model are extracted and semantically annotated based on a knowledge graph, generating a structured set of deformation patterns and corresponding semantic interpretations.

[0009] A bidirectional information transmission channel between the geometric domain and the knowledge domain is constructed based on a 3D mesh model and a set of deformation modes, generating a knowledge-geometric collaborative measurement graph.

[0010] Based on a multi-scale spatiotemporal memory structure, self-evolutionary learning and resource optimization allocation are realized, outputting an adaptive resource allocation strategy and an updated monitoring knowledge base;

[0011] Risk assessments are performed based on the set of transformation patterns, semantic interpretations, and an updated monitoring knowledge base. Intervention recommendations are generated, and risk-level early warning information and multi-level decision support solutions are output.

[0012] In a preferred embodiment, the steps of performing adaptive acquisition and preprocessing operations include:

[0013] A sampling density allocation matrix is ​​constructed based on the risk area data, wherein the sampling density allocation matrix defines the sampling density value for each spatial location;

[0014] Adaptive density image acquisition is performed based on the sampling density allocation matrix to form an original image set containing views from multiple angles and positions;

[0015] The original image set is then subjected to illumination and quality optimization to generate a standardized image set.

[0016] In a preferred embodiment, the sampling density allocation matrix is ​​calculated using the following formula:

[0017] ;

[0018] in, Represents the sampling density allocation matrix. Based on the sampling density constant, For position The risk level function value at the location, This represents the value of the geometric uncertainty function at that location. and These are the first and second weight parameters, respectively.

[0019] In a preferred embodiment, the knowledge-geometric interaction entropy-guided adaptive 3D reconstruction step includes:

[0020] Extract joint knowledge geometric features to generate two types of feature representations: geometric features and semantic features.

[0021] Calculate the feature importance weight matrix to quantify the contribution of each feature point to the reconstruction process;

[0022] Perform weighted feature point matching and camera parameter estimation to generate a set of camera projection matrices;

[0023] Generate multi-view depth maps and fused point clouds, including the three-dimensional geometry of the foundation pit surface;

[0024] Construct a semantically annotated 3D mesh model, which includes geometric surface and semantic annotation information.

[0025] In a preferred embodiment, the knowledge graph-based variant pattern extraction and semantic annotation steps include:

[0026] Establish a deformation representation diagram to express the geometric and semantic information of the 3D mesh model in the form of a graph structure;

[0027] Extract the deformation features of knowledge constraints and generate a deformation pattern feature set;

[0028] Identify and classify deformation pattern types to obtain a set of discrete deformation pattern categories;

[0029] Generate semantic interpretations of deformation patterns, including deformation type labels, causal explanations, development trend predictions, and risk level assessments.

[0030] In a preferred embodiment, the step of constructing a bidirectional information transfer channel between the geometric domain and the knowledge domain includes:

[0031] The uncertainty distribution of the quantified geometric reconstruction is used to generate a three-dimensional spatial field representing the spatial distribution of the reconstruction accuracy.

[0032] The uncertainty distribution of quantified knowledge reasoning is used to generate a three-dimensional spatial field representing the credibility distribution of pattern recognition.

[0033] Construct a set of trigger mappings from geometry to knowledge:

[0034] ;

[0035] in, This represents the set of trigger mappings from geometry to knowledge. Represents the geometric uncertainty threshold. Indicates the risk level threshold. Point Risk level assessment value at the location;

[0036] Construct a set of knowledge-to-geometry trigger mappings to identify areas where geometric reconstruction accuracy needs to be improved first;

[0037] Generate a knowledge geometric collaborative metric graph that integrates geometric accuracy, semantic importance, and uncertainty.

[0038] In a preferred embodiment, the steps for realizing self-evolutionary learning and resource optimization allocation based on a multi-scale spatiotemporal memory structure include:

[0039] Construct a multi-scale spatiotemporal memory storage structure that includes short-term, medium-term, and long-term memory levels;

[0040] Perform experience accumulation and extraction operations to generate a structured set of experience knowledge;

[0041] Candidate resource allocation schemes are generated and evaluated based on resource constraints and knowledge geometry collaborative metric graphs, and the optimal scheme that maximizes system benefits is selected.

[0042] By updating and iterating the knowledge base, the system's cognitive capabilities can be continuously improved.

[0043] In a preferred embodiment, the uncertainty distribution of the geometric reconstruction is calculated using the following formula:

[0044] ;

[0045] in, Point Geometric uncertainty value at the location, Point The standard deviation of the geometric reconstruction, Point The average value of the geometric reconstruction.

[0046] In a preferred embodiment, the uncertainty distribution of knowledge reasoning is calculated using the following formula:

[0047] ;

[0048] in, Point The uncertainty value of knowledge at that location, Represents the set of transformation mode categories. Indicates a given point Features In this case, the point belongs to the deformation category. The probability value, Point The feature vector extracted at the location.

[0049] In a preferred embodiment, the deep foundation pit deformation monitoring system based on close-range photogrammetry includes:

[0050] The adaptive data acquisition and preprocessing module is used to receive risk area data from the knowledge graph, perform adaptive acquisition and preprocessing operations, and generate a standardized image set.

[0051] The knowledge geometric interaction entropy reconstruction module is used to combine a standardized image set with domain knowledge constraints to perform adaptive 3D reconstruction and generate a 3D mesh model with semantic annotations.

[0052] The deformation pattern extraction and annotation module is used to extract and semantically annotate the deformation patterns of the 3D mesh model based on knowledge graphs, and generate a structured set of deformation patterns and corresponding semantic interpretations.

[0053] The uncertainty transmission mechanism module is used to construct a two-way information transmission channel between the geometric domain and the knowledge domain based on the 3D mesh model and the set of deformation modes, and to generate a knowledge-geometric collaborative measurement graph.

[0054] The self-evolutionary learning and resource optimization module is used to achieve continuous evolution of system capabilities and optimized resource allocation based on multi-scale spatiotemporal memory structures.

[0055] The risk assessment and decision support module is used to perform risk assessments and generate intervention recommendations based on the set of transformation patterns, semantic interpretation, and monitoring knowledge base, and output risk classification early warning information and multi-level decision support solutions.

[0056] The beneficial effects of this invention are as follows:

[0057] This invention achieves a dual improvement in the accuracy and efficiency of foundation pit deformation monitoring by constructing a knowledge-geometric dual-state symbiotic system. This system organically integrates geometric observation data with engineering knowledge, enabling it to automatically identify, classify, and interpret various deformation patterns. This not only improves the accuracy of foundation pit deformation monitoring from the centimeter level to the millimeter level, but also tracks the development trend of minute deformations, issuing early warnings before deformation reaches dangerous thresholds, significantly enhancing the safety assurance capability of foundation pit engineering.

[0058] The self-evolutionary learning mechanism based on a multi-scale spatiotemporal memory structure enables this system to continuously accumulate experience from historical projects and transform it into structured knowledge for application in new projects. As the system is used over time, the efficiency of monitoring resource allocation and the accuracy of key area identification both improve. Simultaneously, while maintaining high monitoring quality, the overall operating cost is reduced by approximately 35%. It is suitable for long-term safety monitoring of various deep and large foundation pit projects and has significant engineering application value. Attached Figure Description

[0059] Figure 1 This is a flowchart of the method for monitoring deformation of deep foundation pits based on close-range photogrammetry according to the present invention. Detailed Implementation

[0060] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0061] At least one embodiment of the present invention discloses a method for monitoring the deformation of deep foundation pits based on close-range photogrammetry, such as... Figure 1 As shown, it includes the following steps:

[0062] Step 100: Close-range photogrammetry data acquisition and preprocessing based on the dual-state symbiotic framework;

[0063] It receives risk area data from a knowledge graph, performs adaptive acquisition and preprocessing operations, and outputs a standardized image set. Specifically:

[0064] Step 101: Construct a sampling density allocation matrix based on risk area data: Using the risk area data from the existing knowledge graph as input, generate a sampling density allocation matrix through an adaptive sampling density calculation algorithm. This matrix defines each location in the deep foundation pit space. Sampling density values:

[0065]

[0066] in, Indicates position The sampling density value at that location; This represents the basic sampling density constant, which defines the standard sampling density for areas without special risks. Indicates position The risk level function value at a location quantifies the potential risk level at that location; The value of the geometric uncertainty function at that location reflects the degree of uncertainty in the geometric reconstruction at that location; and These are the third and fourth weighting coefficients, respectively, which control the degree of influence of risk level and geometric uncertainty on sampling density.

[0067] Step 102, perform adaptive density image acquisition: allocate matrix according to sampling density. The system controls a multi-view image acquisition device to obtain raw images of the deep foundation pit, forming a raw image set containing views from multiple angles and positions. ,in , , These are the 1st, 2nd, and nth original images, respectively. This represents the total number of images. The original image set contains visual information about the surface of the deep foundation pit, but may have issues such as uneven lighting and insufficient contrast.

[0068] Step 103, perform image illumination and quality optimization: optimize the original image set. The image is input into the image enhancement module, where it undergoes illumination compensation and quality optimization algorithms to eliminate uneven lighting, enhance texture details, correct color distortion, and generate a standardized image set. ,in , , These are the 1st, 2nd, and nth standardized images, respectively. This represents the total number of standardized images. The standardized image set features balanced lighting conditions and enhanced detail, providing high-quality input data for subsequent 3D reconstruction.

[0069] Step 200, adaptive 3D reconstruction guided by knowledge geometric interaction entropy;

[0070] This step involves inputting a standardized image set. By incorporating domain knowledge constraints, a semantically annotated 3D mesh model is output. Specifically:

[0071] Step 201, Extract joint knowledge geometric features: Standardize the image set... The input is fed into a joint feature extractor, which generates two types of feature representations in parallel: geometric features and semantic features.

[0072]

[0073] in, This represents a knowledge-geometric joint feature extractor, used to extract geometric and semantic features from an image simultaneously; It represents a set of geometric features extracted from an image, including low-level geometric elements such as feature points, edges, and textures; It represents a set of knowledge features extracted from an image, including high-level semantic elements such as material type and structural component type.

[0074] Step 202, calculate the feature importance weight matrix: based on the extracted joint features For each feature point, the knowledge geometric interaction entropy algorithm is used. Calculate importance weights:

[0075]

[0076] in, Representing feature points The importance weight value quantifies the contribution of this point to the reconstruction process; Representing feature points The probability distribution in the geometric domain reflects the geometry of that point; Representing feature points The probability distribution in the knowledge domain reflects the semantic importance of that point; It is a balancing factor used to adjust the relative importance of geometric and semantic information in weight calculation.

[0077] Step 203, perform weighted feature point matching and camera parameter estimation: combine the feature point set and its weight matrix. The input is fed into a multi-view matching algorithm for weighted feature matching, generating a set of camera projection matrices.

[0078]

[0079]

[0080] in, Indicates the first The projection matrix of each viewpoint defines the mapping relationship from three-dimensional space to two-dimensional image; This represents the total number of projection matrices. This represents the camera intrinsic parameter matrix, which includes internal parameters such as focal length and principal point coordinates. This represents the rotation matrix, describing the camera's orientation; This represents the translation vector, describing the camera's position.

[0081] Step 204, Generate multi-view depth maps and fused point clouds: using camera projection matrix sets and standardized image sets The variational multi-view stereo matching algorithm is applied to generate a set of depth maps for each viewpoint:

[0082] ;

[0083] in, , , They represent the 1st, 2nd, and 3rd respectively. Depth map from multiple perspectives; Indicates the total number of depth maps;

[0084] The depth maps are then converted and merged into a unified 3D point cloud. :

[0085]

[0086] in, This represents the inverse projection function, which projects two-dimensional pixel coordinates... and depth value Through projection matrix Convert to three-dimensional space coordinates ; This represents the fused 3D point cloud, which contains the complete geometric structure of the foundation pit surface.

[0087] Step 205, Construct a 3D mesh model with semantic annotation: Convert the point cloud... and semantic features The input is fed into a knowledge-constrained Poisson surface reconstruction algorithm to generate a 3D mesh model with semantic annotations. :

[0088]

[0089] in, Represents a 3D mesh model, containing a set of vertices. Edge set , noodle set and semantic annotation set ; Represents data items that measure the degree of fit between the model and the point cloud; This indicates a smoothing term, ensuring the continuity of the surface; This represents a knowledge constraint term, ensuring that the reconstruction result conforms to known physical laws and engineering experience; and These are the weighting coefficients, which control the strength of the smoothing constraint and the knowledge constraint, respectively.

[0090] Step 300: Extraction and semantic annotation of deformable patterns based on knowledge graph;

[0091] This step converts the semantically annotated 3D mesh model into a structured set of deformation patterns and corresponding semantic interpretations. Specifically:

[0092] Step 301, Establish Deformation Representation Map: Transform the semantically annotated 3D mesh model Convert to deformation characterization diagram The mathematical representation of this diagram is:

[0093]

[0094] in, A deformation representation diagram is a structured graphical representation. This represents the set of nodes in the graph, where each node... This corresponds to a key point on the 3D model; This represents the set of edges in the graph, where each edge... Indicates the spatial connection relationship between key points; This represents an attribute matrix, which contains the characteristic attributes of nodes and edges, such as geometric deformation, material type, and structure type.

[0095] Step 302, extract the deformation features of knowledge constraints: transform the deformation representation diagram. and domain prior knowledge The input is fed into a knowledge-guided graph convolutional network to generate a deformable pattern feature set. :

[0096]

[0097] in, It represents the feature set of deformation modes, capturing the essential characteristics of foundation pit deformation; This represents a knowledge-guided graph convolutional network function used to extract features from graph structures. It represents a set of prior knowledge, including physical constraints and engineering experience, used to guide the feature extraction process.

[0098] Step 303, Identify and classify deformation mode types: Combine deformation mode feature sets The input is fed into a physically constrained unsupervised clustering algorithm to obtain a set of discrete deformation pattern categories. :

[0099]

[0100] in, This represents a set of deformation pattern categories, including various typical deformation patterns; This represents a clustering function used to group similar features; This represents the set of clustering constraints, including physical rationality constraints, spatiotemporal continuity constraints, etc. Indicates the first Each deformation mode category includes multi-dimensional information such as spatial distribution map, deformation degree quantification value, and deformation rate time series.

[0101] Step 304, Generate semantic interpretation of the morphing patterns: This involves creating a set of morphing pattern categories. The input is fed into the knowledge reasoning engine to generate a corresponding set of variant descriptions. :

[0102]

[0103] in, This represents a set of deformation descriptions, containing professional explanations of various deformation modes; This represents a semantic interpretation function that transforms numerical features into understandable textual descriptions. This represents a knowledge base containing professional knowledge such as the causal relationships of deformation types and the mapping of risk levels of deformation characteristics. Indicates the first The semantic description of each deformation pattern includes deformation type labels, possible cause explanations, development trend predictions, and risk level assessments.

[0104] Step 400: Construction of bidirectional information flow for uncertainty propagation mechanism;

[0105] This step, based on the 3D model and deformation pattern data, constructs a bidirectional information transfer channel between the geometric domain and the knowledge domain, outputting a knowledge-geometric co-metric graph. Details are as follows:

[0106] Step 401, Quantify the uncertainty distribution of geometric reconstruction: quantify the 3D mesh model. and raw point cloud data As input, statistical analysis methods are used to calculate the value of each point in space. Geometric reconstruction uncertainty value :

[0107]

[0108] in, Point Geometric uncertainty value at the location, Point The standard deviation of the geometric reconstruction, Point The average value of the geometric reconstruction. The generated geometric uncertainty distribution. It is a three-dimensional spatial field that represents the spatial distribution of reconstruction accuracy within the entire foundation pit area.

[0109] Step 402, Quantify the uncertainty distribution of knowledge reasoning: transform the feature set of the pattern and the set of transformation mode categories As input, each point in the computation space Uncertainty value of knowledge reasoning :

[0110]

[0111] in, Point The uncertainty value of knowledge at that location, Indicates a given point Features In this case, the point belongs to the deformation category. The probability value, Indicates the deformation category, Represents the set of transformation mode categories. Point Features at the location. Distribution of uncertainty in the generated knowledge. Similarly, it is a three-dimensional spatial field that represents the credibility distribution of pattern recognition and semantic understanding within the entire foundation pit area.

[0112] Step 403, construct a set of triggering mappings from geometry to knowledge: combining geometric uncertainty distribution and risk level distribution Determine the set of spatial points whose geometric information should trigger knowledge updates. :

[0113]

[0114] in, This represents the set of trigger mappings from geometry to knowledge. Represents the geometric uncertainty threshold. Indicates the risk level threshold. Point The risk level value at the location, This represents points in space. This mapping set identifies which areas in the foundation pit, due to high geometric uncertainty and high risk level, require priority for knowledge updates and semantic interpretation.

[0115] Step 404, construct a set of trigger mappings from knowledge to geometry: combining knowledge uncertainty distribution and the gradient distribution of knowledge uncertainty Determine the set of spatial points that should trigger geometric precision enhancement based on knowledge information. :

[0116]

[0117] in, This represents the set of trigger mappings from knowledge to geometry. This represents the threshold of knowledge uncertainty. The threshold representing the gradient of knowledge uncertainty. Point Spatial gradient value for knowledge uncertainty This represents points in space. This mapping set identifies which areas in the foundation pit require priority for improving geometric reconstruction accuracy due to high knowledge uncertainty and rapid changes.

[0118] Step 405, Generate a knowledge geometric collaborative metric graph: based on geometric accuracy indicators Knowledge Relevance Index and its corresponding uncertainty distribution and Calculate the knowledge geometric co-metric value in the entire three-dimensional space. :

[0119]

[0120] in, Point Knowledge geometry collaborative metric value at the location, Point Geometric accuracy index value at the location, Point The knowledge relevance index value at the location, and These are the weighting coefficients. Represents a point in space. Point Geometric uncertainty value at the location, Point The knowledge uncertainty value at the given location. The generated knowledge geometric collaborative metric graph. It is a three-dimensional scalar field that integrates geometric precision, semantic importance, and their respective uncertainties, providing a unified reference for system resource allocation.

[0121] Step 500: Self-evolutionary learning and resource optimization allocation based on memory enhancement;

[0122] This step takes into account the knowledge geometry collaborative metric graph, historical deformation data, and system resource constraints to optimize system resource allocation and enhance cognitive capabilities, outputting an adaptive resource allocation strategy and an updated monitoring knowledge base. Details are as follows:

[0123] Step 501, Construct a multi-scale spatiotemporal memory storage structure: Store historical deformed data Co-metric graph with current knowledge geometry Input is fed into the memory storage module to construct a hierarchical spatiotemporal memory structure. :

[0124]

[0125] in, This represents a collection of multi-scale spatiotemporal memory structures, including short-term memory. (Stores high-precision raw data from the last 7 days), Intermediate memory (Storing the main deformation patterns and statistical characteristics of the last 790 days) and long-term memory (Stores abstract patterns and empirical knowledge for over 90 days). This memory structure enables data organization at different time scales and levels of abstraction, allowing the system to efficiently store and access historical information.

[0126] Step 502, perform experience accumulation and extraction operations: transfer the multi-scale spatiotemporal memory structure The input is fed into the experience extraction module, where a pattern mining algorithm generates an experience knowledge set. :

[0127]

[0128] in, This represents an experience knowledge record, which includes a schema description. (Formal representation of the transformation pattern), environmental conditions (Description of environmental conditions that trigger this mode), time characteristics (The temporal evolution characteristics of this pattern) and frequency statistics (The frequency and probability distribution of this pattern) An index representing the experiential knowledge records. The generated set of experiential knowledge. The regularities hidden in historical data are extracted in a structured form to provide an empirical basis for subsequent optimization decisions.

[0129] Step 503, Generate a set of candidate allocation schemes based on resource constraints: This involves considering the system resource constraints... Knowledge Geometric Collaborative Measurement Graph and experience knowledge set The input is fed into the resource allocation optimization module, which generates a set of candidate resource allocation schemes.

[0130]

[0131] in, , , They represent the 1st, 2nd, and 3rd respectively. One candidate resource allocation scheme; This represents the total number of candidate resource allocation schemes;

[0132]

[0133] in, Indicates the first One candidate resource allocation scheme Indicates spatial location Allocated resources , This indicates the total resource constraint. This represents the total number of positions after spatial discretization. Index representing spatial location, Indicates the index of the candidate solution. Each candidate solution It is a feasible resource allocation strategy that satisfies the condition that the total resources do not exceed the constraint.

[0134] Step 504, Simulation Evaluation and Optimal Solution Selection: Set up candidate resource allocation schemes The input is fed into the performance simulation and evaluation module, combined with the knowledge geometry collaborative metric graph. and experience knowledge set Calculate the expected return value for each option. :

[0135]

[0136] in, Representation scheme Expected return value, This represents the resource benefit conversion function. Represents location based on empirical knowledge Importance evaluation function Indicates position Knowledge geometry collaborative metric value at the location, This represents the total number of positions after spatial discretization. Index representing spatial location, This represents the index of the candidate solutions. The optimal resource allocation solution is selected by comparing the expected returns of each solution. :

[0137]

[0138] Optimal resource allocation scheme It clearly specifies how the system should allocate limited computing, storage, and network resources to maximize overall monitoring benefits.

[0139] Step 505, perform knowledge base update and evolution iteration: update the current set of transformation modes. Deformation description set and experience knowledge set Input is fed into the knowledge base update module to generate the updated monitoring knowledge base. :

[0140]

[0141] in, This represents the knowledge base update function. This refers to the existing monitoring knowledge base. This indicates the updated knowledge base. Represents the current set of deformation modes. Represents a set of deformation descriptions. Represents a set of experiential knowledge. The updated knowledge base. It not only includes updated causal relationships of deformation patterns and risk assessment rules, but also newly discovered deformation patterns and the system's adaptive experience to different situations, reflecting the continuous improvement of the system's cognitive ability.

[0142] Step 600: Risk assessment and proactive intervention decision support;

[0143] This step takes as input the updated monitoring knowledge base, the current set of deformation patterns, and resource allocation strategies, performs a risk assessment and generates intervention recommendations, and outputs risk classification early warning information and multi-level decision support solutions. Details are as follows:

[0144] Step 601, Perform dynamic risk threshold calculation: Update the monitoring knowledge base. and the current set of deformation modes The input is fed into the risk threshold calculation module to generate an adaptive risk threshold matrix. :

[0145]

[0146] in, Representation type Deformation in the region The risk threshold, Represents the basic threshold constant. This represents the function for calculating the impact factor, taking into account empirical rules in the knowledge base, current deformation mode characteristics, deformation type, and regional attributes. Index representing the transformation type, Index representing a region This indicates the updated monitoring knowledge base. Represents the current set of deformation modes. The generated risk threshold matrix. This provides a criterion for judging environmental adaptability in subsequent risk assessments.

[0147] Step 602, Generate a multi-dimensional risk scoring map: This involves combining the current deformation mode set... Deformation description set and risk threshold matrix Input into the risk scoring module to calculate a multi-dimensional risk scoring graph. :

[0148]

[0149] in, Indicates the first Risk score, Indicates the weighting coefficient. This represents the risk calculation function. Indicates the first One transformation mode, Indicates the first A variation description, Representation type Deformation in the region The risk threshold, Index representing the transformation type, Index representing a region Indexes representing risk dimensions. Multi-dimensional risk scoring graph. The scoring system includes multiple dimensions such as deformation rate risk, cumulative deformation risk, trend risk, and anomaly risk, comprehensively characterizing the risk status of foundation pit deformation.

[0150] Step 603, Construct a causal risk transmission network: This involves integrating the risk scoring map... and monitoring knowledge base Inputting into the risk propagation model to construct a causal risk propagation network :

[0151]

[0152] in, This indicates a causal risk transmission network. This represents a set of nodes (corresponding to different risk events). This represents a set of edges (corresponding to the causal relationships between risk events). Indicates risk from the event Spread to the event The conditional probability, This represents the function for calculating the probability of risk propagation. This indicates the updated monitoring knowledge base. Represents a risk scoring graph. and Indicates risk event nodes. Causal risk propagation network. Formal representation of the relationships between different risk events enables the system to predict the possible propagation paths and scope of risks.

[0153] Step 604, Generate tiered early warning information and spatial distribution map: This involves creating a multi-dimensional risk scoring map. and causal risk transmission networks Input is fed into the early warning generation module to generate tiered early warning information. and spatial distribution map of early warning :

[0154]

[0155] in, , Indicates a risk level of Set of spatial points ( These correspond to four levels: safe, caution, warning, and danger, respectively. This represents the risk level determination function. Represents a risk scoring graph. This indicates a causal risk transmission network. Represents spatial coordinates. Early warning spatial distribution map. It is a three-dimensional color-coded map that visually displays the risk distribution of the entire foundation pit area, making the monitoring results visual and easy to understand.

[0156] Step 605, Generate a multi-level decision support scheme: Integrate tiered early warning information. Early warning spatial distribution map and monitoring knowledge base Input is fed into the decision support module to generate a multi-level decision support solution. :

[0157]

[0158] in, This indicates an immediate intervention plan (including safety measures that need to be taken immediately). This indicates a short-term intervention plan (including reinforcement measures to be completed within 17 days). This indicates a long-term intervention plan (including recommendations for adjusting monitoring strategies and optimizing the protection system). , and These represent the generating functions for immediate, short-term, and long-term scenarios, respectively. This indicates tiered early warning information. This represents a spatial distribution map of the early warning system. This represents the updated monitoring knowledge base. Multi-level decision support solution. It provides specific operational recommendations for engineering decisions at different time scales, ensuring engineering safety.

[0159] In one embodiment of the present invention, an application example of the aforementioned method for monitoring deformation of deep foundation pits based on close-range photogrammetry is provided:

[0160] This example focuses on a foundation pit project for a super-large underground complex, which has the following characteristics:

[0161] Excavation pit area: approximately 68,500 square meters;

[0162] Excavation pit depth: The deepest point reaches 38.6 meters;

[0163] Surrounding environment: It is adjacent to the Metro Line 13 tunnel to the north (about 23 meters away), a historical building complex to the west (about 35 meters away), and a high-rise office building complex to the east (about 40 meters away).

[0164] Geological conditions: The upper part consists of a fill layer (35 meters) and a sandy silt layer (812 meters), the middle part consists of a silty clay layer (1015 meters), and the lower part consists of moderately weathered granite;

[0165] Groundwater conditions: The shallow groundwater level is buried at a depth of about 57 meters, and the confined water level is buried at a depth of about 2530 meters.

[0166] Support method: A combined support structure of diaphragm wall + three steel pipe supports + prestressed anchor cables is adopted;

[0167] The main technical challenges facing the project include: a sensitive surrounding environment with strict requirements for deformation control; complex geological conditions with uneven weathered rock layers and fault zones; a long construction period (estimated at 24 months) requiring long-term stable monitoring; and variable meteorological conditions with significant seasonal impacts.

[0168] The project adopted this implementation method to construct a self-evolving foundation pit cognitive system with knowledge geometry dual-state symbiosis. The deployment and configuration of the monitoring system are shown in Tables 1 and 2:

[0169] Table 1: Hardware Configuration Table of Monitoring System

[0170]

[0171] Table 2: Distribution of Risk Zones in Excavation Pit

[0172]

[0173] In the early stages of project implementation, the system automatically generated a sampling density allocation matrix based on the risk area distribution and level in Table 2. Taking the RA01 area on the north side of the foundation pit, close to the subway tunnel, as an example, this area has the highest risk level (9 points), and the system allocated a higher sampling density to it than to other areas. Table 3 shows a comparison of the sampling densities for different risk areas:

[0174] Table 3: Examples of Adaptive Sampling Density Allocation in Risk Areas

[0175]

[0176] The sampling density allocation matrix guides the dynamic scheduling of cameras. For example, the system performs 24 high-density image acquisitions per day for area RA01, while only 4 image acquisitions are performed for general areas with lower risk levels. This risk-level-based differentiated acquisition strategy ensures efficient use of limited monitoring resources.

[0177] During the 3D reconstruction process, the system applies the knowledge geometric interaction entropy algorithm to calculate the weights of feature points. Taking a reconstruction process on July 15, 2023 as an example, the system detected a significant increase in geometric uncertainty in region RA05 (the intersection of fault zones in the middle of the foundation pit). The feature point weight matrix calculated by the knowledge geometric interaction entropy algorithm improved the reconstruction accuracy of this region. Table 4 shows the feature point weight distribution in different regions on that day:

[0178] Table 4: Example of feature point weight allocation based on knowledge geometric interaction entropy (July 15, 2023)

[0179]

[0180] This example clearly demonstrates the dynamic control capability of knowledge geometric interaction entropy on the reconstruction process. When the system detects the geometric domain probability of region RA05... and knowledge domain probability When the values ​​are lower (0.45 and 0.25 respectively), it indicates that the uncertainty in that region is high. The system automatically increases the weight value of that region (3.65), thereby improving the reconstruction accuracy of that region by 285%. In contrast, the weight value of general regions is only 0.18 due to their lower uncertainty, resulting in an improvement in reconstruction accuracy of only 12%.

[0181] Through adaptive weight allocation, the system has achieved a refined monitoring strategy that prioritizes key areas under limited resources, thus establishing a positive correlation between reconstruction accuracy and risk level.

[0182] During long-term monitoring, the system has gradually established a rich deformation pattern library. Taking the monitoring of area RA04 (the corner of the south foundation pit) from August 10 to August 30, 2023 as an example, the system successfully identified and classified a potential risk deformation pattern by extracting deformation features from the deformation characterization diagram structure and knowledge constraints, as shown in Table 5:

[0183] Table 5: Examples of Deformation Pattern Extraction and Semantic Annotation in the RA04 Region

[0184]

[0185] This example demonstrates that, without pre-setting specific corner instability modes, the system successfully identified local corner instability risk patterns through graph structure representation and knowledge-constrained feature extraction, and provided accurate semantic interpretation and risk assessment. The system not only identified the abnormal growth trend of deformation but also correlated it with multi-dimensional information such as support axial force, anchor cable prestress, and surrounding surface settlement, thus providing a comprehensive assessment of potential risks.

[0186] One of the core innovations of this system is the establishment of a two-way information transmission mechanism between the geometric domain and the knowledge domain. Taking the system's monitoring of area RA08 (the weak interlayer area at the bottom of the foundation pit) on September 5, 2023 as an example, the system detected abnormal changes in geometric uncertainty and knowledge uncertainty in this area, triggering the two-way information transmission mechanism, as shown in Tables 6 and 7:

[0187] Table 6: Examples of Uncertainty Transmission Mechanisms in the RA08 Region (September 5, 2023)

[0188]

[0189] Table 7: Comparison of the Synergistic Improvement Effects of Reconstruction Accuracy and Risk Perception

[0190]

[0191] Data analysis shows that this implementation method improves reconstruction accuracy by an average of 259% in high-risk areas (RA01, RA05, RA04), while only improving it by 5% in ordinary areas. Simultaneously, the risk identification accuracy improves by an average of 18.9 percentage points in high-risk areas, but only by 4.9 percentage points in ordinary areas. This fully verifies that this implementation method can dynamically adjust system resource allocation according to risk level, achieving a synergistic improvement in accuracy and risk.

[0192] Furthermore, by comparing the system resource allocation at different times, it was found that this implementation automatically allocates most of the computing and storage resources to high-risk areas:

[0193] Computing resource allocation: High-risk areas (accounting for 21% of the total area) received 82.7% of the computing resources;

[0194] Storage resource allocation: Data storage accuracy in high-risk areas is 3.5 times higher than in ordinary areas;

[0195] Detection frequency allocation: The monitoring frequency in high-risk areas is 6 times higher than that in ordinary areas.

[0196] To verify the system's self-evolution capability, the project team conducted a system performance evaluation at different time points (initial deployment, 3 months later, 6 months later, and 12 months later). Specific evaluation data is shown in Table 8.

[0197] Table 8: Validation data on the system's self-evolution capability over time

[0198]

[0199] Compared to traditional systems deployed at the same time (using fixed parameters and algorithms), the performance metrics of traditional systems remained basically unchanged or decreased slightly, as shown in Table 9:

[0200] Table 9: Performance Comparison of Traditional System and This Implementation Method over 12 Months

[0201]

[0202] The above data fully demonstrates that the self-evolution mechanism of this implementation method enables the system's capabilities to continuously improve over time, with an average improvement of over 150% in key performance indicators within one year. In particular, in terms of deformation pattern recognition capabilities, the system autonomously discovered and verified 26 new deformation patterns. These patterns played an important role in later monitoring, and the early identification of 5 key deformation patterns helped the project avoid potential engineering risks and accidents.

[0203] In one embodiment of the present invention, in order to adapt to the application scenario of large-scale foundation pit engineering with a wide monitoring area, numerous monitoring points, and complex and variable environment, and to improve the coverage and collaborative efficiency of the monitoring system, the method for monitoring the deformation of deep and large foundation pits based on close-range photogrammetry further includes:

[0204] Step 10: Construct a multi-level asynchronous decision-making architecture;

[0205] The first step is to perform the initialization steps of the deep foundation pit deformation monitoring method based on close-range photogrammetry, and then construct a three-level computing architecture.

[0206] Specifically, this step receives system topology data, computing node resource information, and a description of collaborative task requirements as input data. Through a hierarchical time-scale mapping algorithm, it generates a multi-level asynchronous decision-making architecture, including:

[0207] Cloud node configuration: a list of central server nodes and their computing power parameters, responsible for executing global planning tasks with an update cycle of 1030 seconds;

[0208] Fog layer node configuration: a list of regional edge nodes and their computing power parameters, responsible for performing regional coordination tasks with a 13-second update cycle;

[0209] Edge layer node configuration: List of robot body nodes and their computing power parameters, responsible for executing local reaction tasks with an update cycle of 50200 milliseconds;

[0210] Inter-layer communication link table: Defines the communication topology and bandwidth parameters between nodes at each layer;

[0211] Decision-making authority allocation table: Specifies the initial decision-making scope and authority limits for each level of node;

[0212] The output multi-level asynchronous decision architecture configuration includes the functional positioning, communication relationships and initial resource allocation of each node in the system, providing basic architectural support for the execution of subsequent asynchronous decisions.

[0213] Step 20: Decompose and allocate time-sensitive tasks;

[0214] This step utilizes a time-sensitive task decomposition algorithm to decompose the overall collaborative task and allocate it to the appropriate decision-making level.

[0215] Input data includes: a list of collaborative task descriptions containing all collaborative tasks to be executed and their attributes; a set of task time parameters containing time constraints such as deadlines and expected execution durations for each task; and a computing resource status table containing the available computing resources and load status of each computing node. Step 31 outputs the multi-level asynchronous decision architecture configuration.

[0216] The time-sensitive task decomposition algorithm calculates the time sensitivity index for each task using the following formula:

[0217]

[0218] in, Indicates task The time sensitivity index, the higher the value, the more sensitive it is to time constraints; Indicates task The deadline, in seconds; This represents the gradient of the impact of response time on task success rate, characterizing the change in task success rate caused by each unit change in response time. This indicates the strength of inter-task dependencies, measuring the degree to which a task is depended upon by other tasks. , , These are weighting coefficients used to adjust the relative importance of each factor.

[0219] Based on the calculated time sensitivity index, the algorithm generates a hierarchical task allocation table, which includes: task hierarchy mapping relationship: specifying at which decision level each task is executed; task update frequency configuration: specifying the decision update cycle for each task; and task resource requirement list: specifying the amount of computing and communication resources required for each task.

[0220] The output hierarchical task allocation table provides a detailed execution plan for collaborative tasks in a three-tier computing architecture, offering task-level guidance for implementing asynchronous decisions.

[0221] Step 30: Generate inter-layer information transmission strategies and decision coordination plans;

[0222] This step integrates an information transmission algorithm based on information entropy and a multi-timescale decision coordination algorithm to ensure system consistency in an asynchronous decision-making environment.

[0223] Input data includes: multi-level node status dataset: containing the current computing and communication status of each node at each level; communication resource allocation matrix: recording the available communication resources in the system and their distribution; historical decision record database: storing past decisions and their effects; multi-level asynchronous decision architecture configuration output by step 31; and hierarchical task allocation table output by step 32.

[0224] Information transfer algorithms based on information entropy are used to calculate the value index of inter-layer information transfer.

[0225]

[0226] in, Indicates a point in time Information From node Passed to node Value indicators; Information Information entropy quantifies the degree to which the uncertainty of information is reduced; This represents the expected increase in the decision-making utility of the receiving node after information transmission. Indicates a point in time Transmitting information Communication costs; Indicates the current time point.

[0227] Multi-timescale decision coordination algorithms calculate the degree of coordination between different decisions:

[0228]

[0229] in, Indicate decision and decision The degree of coordination between them; the higher the value, the better the coordination. Indicates the first A similarity metric function to assess the degree of similarity between decisions in different aspects; Indicates the first A conflict measurement function to assess the degree of conflict in decision-making across different aspects; Indicates the first Weights for various similarity measures; The overall weighting coefficients representing conflicting factors; and These represent the number of types of similarity measures and conflict measures, respectively.

[0230] Combining the calculation results of the two algorithms above, the final output of this step is:

[0231] Inter-layer information transfer strategy table: contains the following: Information transfer triggering condition rule set: defines when information needs to be transferred between layers; Information compression encoding scheme: defines how to compress the transferred information to reduce communication costs; Information transfer priority list: defines the priority order for transferring different types of information;

[0232] The decision coordination scheme includes the following: a decision execution sequence table, which specifies the execution order of decisions at different levels and nodes; a conflict resolution rule set, which defines the handling methods when decision conflicts occur; and a collaborative behavior pattern library, which defines the standard behavior patterns for collaborative decision-making between nodes. The output inter-layer information transmission strategy table and the decision coordination scheme together constitute the key tools for maintaining system consistency in an asynchronous decision-making environment, ensuring the coordination and consistency of distributed nodes under limited information exchange conditions.

[0233] Step 40: Calculate the dynamic resource allocation table and conflict resolution strategy;

[0234] This step applies load-aware task deployment algorithms and resource conflict detection algorithms to achieve dynamic optimization of system resource allocation and prevention and handling of potential conflicts.

[0235] Input data includes: Real-time status report of computing nodes: containing real-time operating indicators such as CPU utilization, memory usage, and energy consumption of each computing node; Decision sequence data stream: containing a list of decisions currently being executed and planned to be executed by each level of nodes; Current status of resource allocation: describing the current allocation of various resources in the system; Environmental complexity assessment results: characterizing the complexity and trend of the current environmental condition; Step 33 outputs the inter-layer information transmission strategy table and decision coordination scheme.

[0236] The load-aware task deployment algorithm calculates task deployment decision metrics:

[0237]

[0238] in, Indicates a point in time The task Deploy to nodes The overall score indicates that the deployment is more reasonable; the higher the score, the more reasonable the deployment. Indicates at node Execute the task The calculated energy consumption is expressed in joules. Indicates the task Deploy to nodes The energy consumption generated during communication is measured in joules. Indicates at node Execute the task The delay time, in milliseconds; Represents a node At the point of time The resource utilization rate is between 0 and 1; , , , These are weighting coefficients that are dynamically adjusted based on the system state.

[0239] Resource conflict detection algorithms identify potential decision conflicts:

[0240]

[0241] in, Indicate decision and decision Whether there is a resource conflict is indicated by 1, where 0 indicates that there is no conflict. Indicate decision Resources The set of uses includes the amount of resources used and the time interval of use; Indicate decision Resources The set of uses includes the amount of resources used and the time interval of use; Indicate decision The spatial influence area refers to the physical spatial range affected by the decision. Indicate decision The spatial influence area refers to the physical spatial range affected by the decision. It represents the set of all resources in the system.

[0242] Based on the calculation results of the above algorithm, this step generates the following output:

[0243] Dynamic Resource Allocation Table: Contains the following: Task Migration Instruction Sequence: Specifies which tasks need to be migrated to which nodes; Resource Reallocation Scheme: Describes the optimal allocation method for system resources; Temporary Transfer List of Decision Authority: Specifies the rules for the temporary transfer of decision authority under specific conditions;

[0244] Conflict Resolution Strategy Document: Includes the following: Decision Priority Adjustment List: Specifies the priority order among conflict decisions; Resource Competition Mediation Scheme: Details how to resolve resource competition issues; Spatial Impact Area Coordination Rules: Specifies the operational coordination methods for overlapping impact areas;

[0245] The output dynamic resource allocation table and conflict resolution strategy document together ensure the efficient use of system resources and minimize conflicts in decision execution, enabling the asynchronous decision-making framework to remain efficient and stable in actual operation.

[0246] Step 50: Generate an adaptive adjustment parameter set and performance optimization scheme;

[0247] This step combines an adaptive time step control algorithm and a decision performance evaluation algorithm to achieve continuous optimization and adjustment of system parameters.

[0248] Input data includes: Execution efficiency monitoring dataset: contains efficiency indicators for decision execution at each level, such as decision accuracy and execution latency; Environmental change rate record: records the rate and trend of change of various parameters in the environment; Historical decision effect database: stores the execution effect and performance data of historical decisions; Resource utilization time series: records the time changes in system resource utilization; Step 40 outputs a dynamic resource allocation table and conflict resolution strategy document.

[0249] The adaptive time step control algorithm calculates the optimal update period for decisions at each level:

[0250]

[0251] in, Indicates the first The decision update cycle for the next time point at each level, in milliseconds; Indicates the first The decision update cycle for the hierarchy at the current point in time, in milliseconds; Indicates the first The current execution efficiency of the level, with a normalized value between 0 and 1; This represents the target execution efficiency, which is usually a preset optimal efficiency value; It represents the rate of change of environmental complexity, characterizing the speed of environmental change; and This is an adjustment factor used to balance the impact of execution efficiency and environmental changes.

[0252] The decision performance evaluation algorithm calculates the overall system performance score:

[0253]

[0254] in, Indicates the system at a given time point. The overall performance score indicates that the higher the score, the better the performance. Indicates the number of tasks completed; Indicates the total number of tasks; Indicates the total amount of available resources; Indicates the total amount of resources used; This represents the reference delay time, which is typically the ideal delay for the system design. This represents the average delay time measured in practice. Indicates the number of conflicting decisions; Indicates the total number of decisions; , , , These are the weighting coefficients for each indicator.

[0255] Based on performance scoring, the system optimization objective can be expressed as:

[0256]

[0257] in, This represents the set of adjustable parameters of the system, including the weight coefficients and thresholds of each algorithm; Indicates the deadline. Historical data; Indicates in the parameter and historical data Under the conditions, future time points Expected performance score; This indicates the time interval for prediction, that is, the length of time from the current point in time to the future.

[0258] Based on the calculation results of the above algorithm, this step generates the following output:

[0259] Adaptive time step parameter set: includes the following: decision update cycle table for each level: specifies the optimal decision update frequency for each level; time step adjustment trigger condition: specifies when the decision update cycle needs to be adjusted; dynamic allocation instruction for computing resources: specifies how to adjust computing resources according to the decision frequency;

[0260] System performance optimization plan: includes the following: Algorithm parameter adjustment list: specifies the parameter values ​​that need to be adjusted in each algorithm; Decision boundary fine-tuning rules: specifies how to adjust the boundaries between decision levels; Resource allocation optimization strategy: details the optimization direction of resource allocation;

[0261] The output adaptive time step parameter set and system performance optimization scheme together realize the continuous self-optimization of the system, enabling the asynchronous decision-making framework to adapt to the dynamic adjustment of environmental changes and task requirements, and maintain long-term high-efficiency operation.

[0262] This implementation achieves the following specific technical effects through a multi-level asynchronous decision-making framework and a three-level edge-fog-cloud computing architecture:

[0263] Distributing computational load and reducing communication overhead;

[0264] Computational load distribution effect: Compared with the traditional centralized architecture, the computing load of the central server is reduced, and computing tasks are reasonably distributed on nodes at different levels;

[0265] The degree of reduction in communication bandwidth: The overall communication bandwidth requirement of the system is reduced, and the information entropy-based transmission algorithm reduces the amount of unnecessary information transmitted;

[0266] Improved communication latency: Average communication latency between nodes is reduced; Real-time response capability is improved;

[0267] Reduced response time: The average response time of the system to environmental changes has been reduced from seconds to milliseconds;

[0268] Decision-making and execution efficiency: Emergency task handling capabilities have been improved, achieving millisecond-level local response capabilities;

[0269] Processing throughput: The increase in the number of decisions the system can process per second;

[0270] Optimize the balance between coordination and autonomy;

[0271] Reduced coordination overhead: Inter-level coordination and communication overhead is reduced while maintaining overall system consistency;

[0272] Decision consistency index: Improved coordination and consistency at key decision points ensures global consistency of system behavior;

[0273] Increased local autonomy: The autonomous decision-making ability of edge nodes is improved, reducing the decision-making pressure on central nodes;

[0274] Enhanced system scalability and robustness;

[0275] Scalability: The system can seamlessly scale up to support hundreds of robots, instead of dozens.

[0276] Fault tolerance: When 50% of the edge nodes fail randomly, the system can still maintain 85% of its functionality, improving system robustness;

[0277] Abnormal recovery speed: The system recovery time from partial failures is reduced from minutes to seconds;

[0278] Resource utilization efficiency has been comprehensively optimized;

[0279] Computational resource utilization: The average utilization rate of the system's computing resources has been improved, reducing resource idleness;

[0280] Energy consumption reduction: Under the condition of completing the same task, the overall energy consumption of the system is reduced;

[0281] Task completion quality: The success rate of task completion has increased.

[0282] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments based on the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for monitoring deformation of deep foundation pits based on close-range photogrammetry, characterized in that, Includes the following steps: Receive risk area data from the knowledge graph, perform adaptive acquisition and preprocessing operations, and generate a standardized image set; By combining standardized image sets with domain knowledge constraints, adaptive 3D reconstruction guided by knowledge geometric interaction entropy is performed to generate a 3D mesh model with semantic annotations. The deformation patterns of a 3D mesh model are extracted and semantically annotated based on a knowledge graph, generating a structured set of deformation patterns and corresponding semantic interpretations. A bidirectional information transmission channel between the geometric domain and the knowledge domain is constructed based on a 3D mesh model and a set of deformation modes, generating a knowledge-geometric collaborative measurement graph. Based on a multi-scale spatiotemporal memory structure, self-evolutionary learning and resource optimization allocation are realized, outputting an adaptive resource allocation strategy and an updated monitoring knowledge base; Risk assessments are performed based on the set of transformation patterns, semantic interpretations, and an updated monitoring knowledge base. Intervention recommendations are generated, and risk classification and early warning information and multi-level decision support solutions are output. The adaptive 3D reconstruction steps guided by knowledge geometric interaction entropy include: Extract joint knowledge geometric features to generate two types of feature representations: geometric features and semantic features. Calculate the feature importance weight matrix to quantify the contribution of each feature point to the reconstruction process; Perform weighted feature point matching and camera parameter estimation to generate a set of camera projection matrices; Generate multi-view depth maps and fused point clouds, the fused point clouds containing the three-dimensional geometry of the foundation pit surface; Construct a semantically annotated 3D mesh model, including geometric surface and semantic annotation information; The feature importance weight matrix is ​​calculated based on the extracted joint features. For each feature point, the knowledge geometric interaction entropy algorithm is used. Calculate importance weights: ； in, Representing feature points The importance weight value quantifies the contribution of this point to the reconstruction process; Representing feature points The probability distribution in the geometric domain reflects the geometry of that point; Representing feature points The probability distribution in the knowledge domain reflects the semantic importance of that point; It serves as a balancing factor, used to adjust the relative importance of geometric and semantic information in weight calculation; The weighted feature point matching process involves inputting the feature point set and its weight matrix into a multi-view matching algorithm, performing weighted feature matching, and generating a set of camera projection matrices. The generation of multi-view depth maps and fused point clouds is achieved by using a set of camera projection matrices and a normalized image set. The variational multi-view stereo matching algorithm is applied to generate a set of depth maps for each viewpoint: ； in, , , They represent the 1st, 2nd, and 3rd respectively. Depth map from multiple perspectives; Indicates the total number of depth maps; The depth maps are then converted and merged into a unified 3D point cloud. : ； in, This represents the inverse projection function, which projects two-dimensional pixel coordinates... and depth value Through projection matrix Convert to three-dimensional space coordinates ; This represents the fused 3D point cloud, which contains the complete geometric structure of the pit surface; The steps to construct a bidirectional information transfer channel between the geometric domain and the knowledge domain include: The uncertainty distribution of the quantified geometric reconstruction is used to generate a three-dimensional spatial field representing the spatial distribution of the reconstruction accuracy. The uncertainty distribution of quantified knowledge reasoning is used to generate a three-dimensional spatial field representing the credibility distribution of pattern recognition. Construct a set of trigger mappings from geometry to knowledge: ； in, This represents the set of triggering mappings from geometry to knowledge, and represents the geometric uncertainty threshold. Indicates the risk level threshold. Point The risk level assessment value of the location, Point The geometric uncertainty value at the location; Construct a set of knowledge-to-geometry trigger mappings to identify areas where geometric reconstruction accuracy needs to be improved first; Generate a knowledge geometric co-metric graph that integrates geometric accuracy and semantic importance, along with their corresponding geometric uncertainty distribution and knowledge uncertainty distribution; The steps for achieving self-evolutionary learning and optimized resource allocation based on multi-scale spatiotemporal memory structures include: Construct a multi-scale spatiotemporal memory storage structure that includes short-term, medium-term, and long-term memory levels; Perform experience accumulation and extraction operations to generate a structured set of experience knowledge; Based on resource constraints, knowledge geometry collaborative metric graphs, and experience knowledge sets, candidate resource allocation schemes are generated and evaluated, and the optimal scheme that maximizes system benefits is selected. By updating and iterating the knowledge base, the system's cognitive capabilities can be continuously improved.

2. The method for monitoring deformation of deep foundation pits based on close-range photogrammetry according to claim 1, characterized in that, The steps for performing adaptive acquisition and preprocessing operations include: A sampling density allocation matrix is ​​constructed based on the risk area data, wherein the sampling density allocation matrix defines the sampling density value for each spatial location; Adaptive density image acquisition is performed based on the sampling density allocation matrix to form an original image set containing views from multiple angles and positions; The original image set is then subjected to illumination and quality optimization to generate a standardized image set.

3. The method for monitoring deformation of deep foundation pits based on close-range photogrammetry according to claim 2, characterized in that, The sampling density allocation matrix is ​​calculated using the following formula: ； in, Represents the sampling density allocation matrix. Based on the sampling density constant, For position The risk level function value at the location, This represents the value of the geometric uncertainty function at that location. and These are the first and second weight parameters, respectively.

4. The method for monitoring deformation of deep foundation pits based on close-range photogrammetry according to claim 1, characterized in that, The steps of knowledge graph-based pattern extraction and semantic annotation include: Establish a deformation representation diagram to express the geometric and semantic information of the 3D mesh model in the form of a graph structure; Extract the deformation features of knowledge constraints and generate a deformation pattern feature set; Identify and classify deformation pattern types to obtain a set of discrete deformation pattern categories; Generate semantic interpretations of deformation patterns, including deformation type labels, causal explanations, development trend predictions, and risk level assessments.

5. The method for monitoring deformation of deep foundation pits based on close-range photogrammetry according to claim 4, characterized in that, The uncertainty distribution of the geometric reconstruction is calculated using the following formula: ； in, Point Geometric uncertainty value at the location, Point The standard deviation of the geometric reconstruction, Point The average value of the geometric reconstruction.

6. The method for monitoring deformation of deep foundation pits based on close-range photogrammetry according to claim 4, characterized in that, The uncertainty distribution of knowledge reasoning is calculated using the following formula: ； in, Point The uncertainty value of knowledge at that location, Represents the set of transformation mode categories. Indicates a given point Features In this case, the point belongs to the deformation category. The probability value, Point The feature vector extracted at the location.

7. A deep foundation pit deformation monitoring system based on close-range photogrammetry, characterized in that, The steps for performing the deep foundation pit deformation monitoring method based on close-range photogrammetry as described in any one of claims 1-6 include: The adaptive data acquisition and preprocessing module is used to receive risk area data from the knowledge graph, perform adaptive acquisition and preprocessing operations, and generate a standardized image set. The knowledge geometric interaction entropy reconstruction module is used to combine a standardized image set with domain knowledge constraints to perform adaptive 3D reconstruction and generate a 3D mesh model with semantic annotations. The deformation pattern extraction and annotation module is used to extract and semantically annotate the deformation patterns of the 3D mesh model based on knowledge graphs, and generate a structured set of deformation patterns and corresponding semantic interpretations. The uncertainty transmission mechanism module is used to construct a two-way information transmission channel between the geometric domain and the knowledge domain based on the 3D mesh model and the set of deformation modes, and to generate a knowledge-geometric collaborative measurement graph. The self-evolutionary learning and resource optimization module is used to achieve continuous evolution of system capabilities and optimized resource allocation based on multi-scale spatiotemporal memory structures. The risk assessment and decision support module is used to perform risk assessments and generate intervention recommendations based on the set of transformation patterns, semantic interpretation, and monitoring knowledge base, and output risk classification early warning information and multi-level decision support solutions.

Images