A construction engineering high-altitude operation edge protection stability intelligent monitoring method

Abstract

The present application relates to a kind of construction engineering high-altitude operation edge protection stability intelligent monitoring method, to improve the precision and self-adaptability of structure risk early warning in construction phase.The present application proposes a risk identification method based on building information model progress data and field positioning beacon, which integrates construction phase semantic label, disturbance knowledge graph and multi-source measured data.The core of the present application includes: generating construction semantic label by spatial and time multi-dimensional alignment, combining knowledge graph to intelligently retrieve phase dominant disturbance source and physical parameters, using historical monitoring curve aggregation and deep learning model to realize short-term disturbance correction, nonlinearly fusing the two to construct individualized dynamic threshold envelope, based on continuous over-threshold detection to carry out graded early warning and contribution heat analysis, and loop feedback optimization related parameters and knowledge base.The present application can realize high-reliability risk identification and early warning, improve the intelligent level of structure monitoring system and engineering safety guarantee capability for complex working conditions and environmental disturbance.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and risk assessment of structural stability at construction sites, and in particular to an intelligent monitoring method for the stability of edge protection during high-altitude operations in construction projects. Background Technology

[0002] Existing intelligent monitoring and risk assessment technologies for the stability of edge protection structures in high-altitude construction projects generally employ state identification methods centered on fixed threshold judgment and sliding window statistical analysis. One mainstream approach involves setting a set of structural response thresholds or alarm standards early in the monitoring system, such as root mean square acceleration, peak displacement, and bolt force attenuation percentage. Measured data is then compared to these thresholds; exceeding these thresholds triggers an alarm. This method facilitates rapid deployment in engineering projects, is simple to implement, and is suitable for site environments with relatively simple working conditions or predictable disturbance types. Meanwhile, some research introduces Bayesian regression, fuzzy logic rule bases, or simple variational inference mechanisms, attempting to self-adjust alarm limits through data learning, expanding the applicability of thresholds and improving sensitivity to certain dynamic working conditions. More recent, cutting-edge solutions integrate BIM information, utilize reinforcement learning algorithms such as Q-learning, or hybrid Copula distributions to further enhance the modeling capabilities of multi-source heterogeneous data.

[0003] However, existing methods for monitoring and assessing the risks of protective structures near edges during high-altitude operations have the following shortcomings: Fixed thresholds are prone to generating numerous false alarms and missed alarms in actual operations due to their inability to adapt to the dynamic changes in the construction site at different stages, thus reducing the reliability of early warnings. Furthermore, the lack of deep integration with the BIM progress model and on-site construction stage data makes it impossible to accurately identify the dominant disturbances in the current process, resulting in insufficient ability to identify stage-specific risks. While the mathematical model can adaptively adjust the threshold to some extent, the lack of introduction of engineering process semantics and physical constraints leads to a lag in the response of threshold adjustments to complex working conditions, resulting in poor interpretability. Feedback mechanisms are mostly reactive and lack a closed-loop mechanism to achieve self-evolution and precise correction of the threshold model, resulting in weak knowledge reuse capabilities for similar projects. Summary of the Invention

[0004] This application provides an intelligent monitoring method for the stability of edge protection during high-altitude operations in building engineering, aiming to solve one of the problems or issues of the existing technology mentioned in the background.

[0005] This application provides an intelligent monitoring method for the stability of edge protection during high-altitude operations in construction engineering, specifically including: S1: Obtain the progress data stream of the building information model and the on-site positioning beacon data, generate semantic tags for construction stages with spatiotemporal attributes based on time window alignment and spatial coordinate matching, and construct a stage disturbance knowledge graph containing the root node of the working condition, the disturbance source type node and the physical constraint attribute node.

[0006] S2: Based on the semantic tags of the construction stage, retrieve the pre-set stage disturbance knowledge graph, extract the dominant disturbance source type and physical constraint attributes under the current working condition to generate stage disturbance feature vector.

[0007] S3: Utilize the stage disturbance feature vector retrieval and matching to retrieve the parameter time series curves of similar projects with the current working conditions from the historical building engineering monitoring database. After normalization, construct a benchmark stability decay trajectory template that characterizes the structural response law of the construction stage.

[0008] S4: Acquire real-time sensor data streams and calculate vibration energy entropy, displacement trend slope, and temperature and humidity coupling offset, and input these as short-term disturbance input conditions into the long short-term memory network to output residual compensation coefficients.

[0009] S5: Perform nonlinear fusion calculations on the predicted values ​​of the baseline stability decay trajectory template and the residual compensation coefficient to generate a dynamic threshold envelope for the current construction stage.

[0010] S6: Determine whether the measured structural state value within three consecutive sampling periods exceeds the upper or lower bound of the dynamic threshold envelope. If it does, trigger a graded early warning signal and generate a heat map of stage matching degree and disturbance contribution.

[0011] S7: Receive the on-site handling records and threshold execution effect feedback data after the early warning event is closed, and inject the data including stage labels, false alarm and missed alarm indicators and response delay duration into the stage disturbance knowledge graph.

[0012] S8: Based on the injected feedback data, the association weights in the stage disturbance knowledge graph are iteratively updated to optimize the fitting accuracy of the benchmark stability decay trajectory template under the same construction stage in order to complete the adaptive adjustment closed loop.

[0013] The intelligent monitoring method for the stability of edge protection in high-altitude operations in building engineering provided in this application has the following beneficial effects: (1) By deeply integrating the BIM 4D progress model with the on-site RFID / Bluetooth beacon positioning data, a semantic tagging system for construction stages with spatiotemporal coordinates is constructed. For the first time, dynamic perception and structured expression of the evolution of construction procedures are realized, effectively overcoming the inherent defects of static threshold setting and detachment from engineering context in traditional stability monitoring systems. Existing technologies generally rely on fixed thresholds or sliding window statistical methods for anomaly detection, which are difficult to adapt to complex and ever-changing construction site environments and are prone to false alarms due to stage-based normal operation disturbances. However, this solution is based on the "stage-disturbance" knowledge graph linkage mechanism. When each construction stage starts, it automatically matches the stability decay law of edge protection structures in similar historical projects and generates a benchmark trajectory template with representative working conditions. This provides clear technological background support for risk assessment, significantly improves the rationality and scenario adaptability of threshold setting, and effectively solves the technical problem of difficulty in balancing monitoring sensitivity and robustness in multi-working-condition cross environments.

[0014] (2) A lightweight LSTM network is introduced to perform residual modeling on short-term disturbance features such as vibration energy entropy, displacement trend slope, and temperature and humidity coupling offset collected in real time. Dynamic compensation coefficients are output and superimposed onto the baseline trajectory template through a nonlinear fusion mechanism to form a personalized dual-boundary dynamic threshold envelope for the current construction stage. This significantly enhances the system's ability to capture early signs such as sudden local instability and material performance degradation. Compared with traditional methods such as Bayesian regression, Copula joint distribution modeling, or fuzzy logic rule bases, which suffer from slow convergence and poor interpretability when processing high-dimensional heterogeneous monitoring data, this scheme achieves a risk evolution inference path with clear physical meaning through a semantically driven hierarchical modeling architecture. This not only reduces the dependence on manual experience parameter tuning but also significantly improves the response speed and early warning time. It is especially suitable for major engineering construction sites with tight schedules and frequent process changes, providing more timely and interpretable support for safety decisions.

[0015] (3) A closed-loop adaptive optimization mechanism is constructed, in which the execution effect of each early warning event (such as false alarms, false alarms, and response delays) and on-site handling records are back-injected into the stage disturbance knowledge graph, continuously iterating and optimizing the accuracy of the baseline trajectory template for subsequent similar stages, thereby forming a complete intelligent closed loop of "perception-analysis-early warning-feedback-evolution". This mechanism breaks through the bottlenecks of sample sparsity and high training costs faced by reinforcement learning or probabilistic inference methods such as Q-learning and variational inference in actual deployment. It can achieve gradual accumulation and transfer reuse of knowledge without large-scale historical labeled data, significantly enhancing the long-term applicability and generalization ability of the system. Combined with the "stage matching degree" and "disturbance contribution heat map" pushed by the mobile terminal, managers can intuitively identify the difference between real structural risks and normal construction disturbances, greatly improving the efficiency of on-site judgment and the accuracy of handling, and truly realizing the transformation from passive alarm to proactive cognitive safety prevention and control.

[0016] In summary, this solution innovatively constructs an engineering context understanding framework with the ability to "understand processes, recognize working conditions, and know evolution" by deeply embedding construction management semantics into the entire process of monitoring threshold generation. It completely breaks away from the path dependence on purely mathematical fitting technology and achieves substantial breakthroughs in improving monitoring accuracy, reducing false alarm rates, and enhancing system interpretability and adaptability. It provides a new, efficient, intelligent, and sustainably evolving technological paradigm for the stability monitoring of edge protection structures in complex construction environments. Attached Figure Description

[0017] Figure 1 This is the main flowchart of an intelligent monitoring method for the stability of edge protection during high-altitude operations in construction engineering.

[0018] Figure 2 This is a sub-flowchart of an intelligent monitoring method for the stability of edge protection during high-altitude operations in construction engineering.

[0019] Figure 3 This is another sub-flowchart of a method for intelligent monitoring of the stability of edge protection during high-altitude operations in construction engineering.

[0020] Figure 4 This is an application environment diagram of an intelligent monitoring method for the stability of edge protection during high-altitude operations in building engineering, as shown in one embodiment.

[0021] Figure 5 An internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0023] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0024] like Figure 1 As shown, this application provides an intelligent monitoring method for the stability of edge protection during high-altitude operations in construction engineering, specifically including: S1: Obtain the progress data stream of the building information model and the on-site positioning beacon data, generate semantic tags for construction stages with spatiotemporal attributes based on time window alignment and spatial coordinate matching, and construct a stage disturbance knowledge graph containing the root node of the working condition, the disturbance source type node and the physical constraint attribute node.

[0025] S2: Based on the semantic tags of the construction stage, retrieve the pre-set stage disturbance knowledge graph, extract the dominant disturbance source type and physical constraint attributes under the current working condition to generate stage disturbance feature vector.

[0026] S3: Utilize the stage disturbance feature vector retrieval and matching to retrieve the parameter time series curves of similar projects with the current working conditions from the historical building engineering monitoring database. After normalization, construct a benchmark stability decay trajectory template that characterizes the structural response law of the construction stage.

[0027] S4: Acquire real-time sensor data streams and calculate vibration energy entropy, displacement trend slope, and temperature and humidity coupling offset, and input these as short-term disturbance input conditions into the long short-term memory network to output residual compensation coefficients.

[0028] S5: Perform nonlinear fusion calculations on the predicted values ​​of the baseline stability decay trajectory template and the residual compensation coefficient to generate a dynamic threshold envelope for the current construction stage.

[0029] S6: Determine whether the measured structural state value within three consecutive sampling periods exceeds the upper or lower bound of the dynamic threshold envelope. If it does, trigger a graded early warning signal and generate a heat map of stage matching degree and disturbance contribution.

[0030] S7: Receive the on-site handling records and threshold execution effect feedback data after the early warning event is closed, and inject the data including stage labels, false alarm and missed alarm indicators and response delay duration into the stage disturbance knowledge graph.

[0031] S8: Based on the injected feedback data, the association weights in the stage disturbance knowledge graph are iteratively updated to optimize the fitting accuracy of the benchmark stability decay trajectory template under the same construction stage in order to complete the adaptive adjustment closed loop.

[0032] Step S1: Obtain the building information model progress data stream and on-site positioning beacon data; generate construction stage semantic tags carrying spatiotemporal attributes based on time window alignment and spatial coordinate matching; and construct a stage disturbance knowledge graph containing work condition root nodes, disturbance source type nodes, and physical constraint attribute nodes. Specifically, this includes: S1.1: Obtain the task time window parameters and component type attributes from the building information model progress data stream, and collect the terminal coordinate sequence from the site positioning beacon data. Perform discretization time slicing processing on the task time window parameters to generate a standard time grid, and simultaneously perform Kalman filtering denoising operation on the terminal coordinate sequence to output a high-precision spatial positioning point set.

[0033] The process begins by acquiring task time window parameters and component type attributes from the building information model (BIM) schedule data stream as initial inputs. Discretized time slicing based on construction plan granularity is then applied to the task time window parameters, mapping tasks of different durations to time units of uniform length. A standard time grid suitable for multi-source data synchronization is generated using time window boundary alignment rules. When collecting terminal coordinate sequences from on-site positioning beacon data, the weights of the original positioning points are determined using a multipath signal strength and time arrival difference model, forming a coordinate time sequence with confidence markers. Kalman filtering is then applied to this coordinate time sequence to denoise it. Prior estimates used for state prediction are fused with sensor measurements based on the minimum mean square error criterion. Positioning points are continuously corrected using the state transition matrix and process noise covariance matrix to eliminate instantaneous deviations. The filtered positioning point set undergoes spatial redundancy removal and homogenization resampling to reduce interference from high-density noise points in subsequent spatial matching. Through these processing methods, the original task time window parameters and positioning coordinate sequences are transformed into a standard time grid and a high-precision spatial positioning point set that can directly drive spatiotemporal index matching, providing a precise spatiotemporal foundation for semantic tag generation during the construction phase.

[0034] For example, the time window parameter for a certain task is extracted from the building information model progress data stream as 08:00 to 10:30, and the component type attribute is rebar-tied component. Discretization time slicing is performed in 15-minute units to obtain 10 standard time grid units. The terminal coordinate sequence obtained from the on-site positioning beacon has an original sampling frequency of 1Hz and an azimuth accuracy of ±0.8 meters. Through Kalman filtering, the diagonal elements of the process noise covariance matrix are set to 0.04, and the diagonal elements of the observation noise covariance matrix are set to 0.01. With the state transition matrix being an identity matrix, the Kalman gain is calculated and the state vector is updated in each iteration, resulting in a high-precision spatial positioning point set with an azimuth accuracy better than ±0.25 meters. Further resampling evens out this point set to one positioning point per minute, ensuring a one-to-one correspondence with the standard time grid. Verification results show that this processing chain maintains positioning continuity and stability even in construction sites with significant environmental interference, providing accurate input for the subsequent spatiotemporal index matching operation in S1.2.

[0035] S1.2: Based on the standard time grid and the high-precision spatial positioning point set, the spatial distribution of real-time workers and mechanical equipment is mapped to the construction area corresponding to the building information model progress data stream using the spatiotemporal index matching method. The spatial overlap and temporal consistency are calculated to generate a spatiotemporal correlation confidence matrix.

[0036] Based on the standard time grid and high-precision spatial positioning point set generated in step S1.1, the standard time grid is used as the reference coordinate system for time matching, and the high-precision spatial positioning point set is used as the reference coordinate system for spatial matching. These are then input into the initialization module of the spatiotemporal index matching method.

[0037] The real-time collected positioning coordinates of workers and mechanical equipment are divided and mapped to corresponding time grid segments according to timestamps. The spatial index builder is called to organize a high-precision spatial positioning point set in an R-tree structure to achieve fast spatial query and range matching.

[0038] In the spatial matching stage of the spatiotemporal index matching method, the intersection area of ​​the spatial distribution of real-time workers and mechanical equipment with the geometric boundary of the construction task area in the building information model progress data stream is calculated, and the spatial overlap is determined by the ratio of the intersection area to the total area of ​​the reference area.

[0039] In the time matching stage of the spatiotemporal index matching method, the absolute value of the difference between the real-time positioning event timestamp and the center point of the construction task time window is calculated, and normalized based on the preset time error tolerance interval to obtain the time matching degree.

[0040] The spatial overlap and temporal matching degree are weighted and fused in the correlation calculator to obtain the spatiotemporal correlation confidence matrix. The matrix element values ​​represent the matching strength of workers or mechanical equipment in a certain construction task area and time segment. Through the above processing method, the standard time grid and high-precision spatial positioning point set generated in the previous step are transformed into a spatiotemporal correlation confidence matrix that quantifies the degree of spatiotemporal matching, thereby greatly improving the accuracy of spatial and temporal matching in the construction stage.

[0041] For example, in a high-altitude operation project, the standard time grid interval is 5 minutes, and the coordinate accuracy of the high-precision spatial positioning point set is ±5 cm. The positioning data for the workers are (20.35, 15.60) and (21.20, 16.10), and the positioning data for the machinery and equipment is (19.95, 14.80). The boundary area of ​​the construction task area is 200 square meters, and the spatial intersection area is 150 square meters. The spatial overlap is calculated as follows: =0.75. The center point of the construction task time window is 12:00:00, the timestamp of a certain positioning event is 12:02:00, the absolute value of the time difference is 120 seconds, the maximum allowable time difference is 300 seconds, and the time consistency is calculated as follows: =0.6. The spatial overlap (0.75) and temporal consistency (0.6) are weighted at 0.5:0.5 to calculate the spatiotemporal correlation confidence score. =0.675. In the system's confidence matrix, this value is directly used for subsequent determination of the dominant construction process type and generation of stage semantic labels. The verification results show that the matching accuracy of this construction stage is significantly improved, effectively reducing the stage identification error.

[0042] S1.3: Based on the mapping relationship with the highest confidence in the spatiotemporal correlation confidence matrix, extract the construction task name and planned construction period identifier corresponding to the current work area, and perform semantic parsing on the construction task name in combination with the predefined process logic constraint rule library to determine the current dominant construction process type.

[0043] The input conditions include the previously generated spatiotemporal correlation confidence matrix and construction task metadata from the building information model (BIM) schedule data stream. A maximum value operation is performed on the spatiotemporal correlation confidence matrix, and the index position corresponding to the mapping relationship with the highest confidence is used as the index basis for extracting construction tasks in the current work area. Based on this index, the construction task name field and planned duration identifier field are retrieved from the BIM schedule data stream, ensuring consistency between the index position and the data stream time window to avoid misaligned extraction. The extracted construction task name is then used to call the process logic constraint rule base. During the call, the construction task name is segmented, and the segmentation results are mapped to the process keyword set in the rule base to generate a matching degree distribution vector. A maximum value determination is performed on the matching degree distribution vector, and the process type associated with the corresponding keyword is used as the candidate value for the current dominant construction process. This is combined with contextual logic constraints in the rule base (such as process dependencies and resource usage conflict patterns) for validity verification. The process type that passes the validity verification is used as the final determination result, outputting the dominant construction process type data required for subsequent construction phase semantic tagging. By using matrix maximum value location, metadata retrieval, and rule base matching and verification, the spatiotemporal correlation results of the previous step are transformed into the dominant construction process type for semantic determination in the current stage, thereby achieving precise process determination and stage label generation driven by the process.

[0044] For example, in the scenario of concrete pouring on the 15th floor of a high-rise building, the spatiotemporal correlation confidence matrix is ​​of size 20×15, with a maximum value of The corresponding index position (row 8, column 12) corresponds to the construction task name "Concrete Pouring - Main Structure" in the Building Information Model (BIM) schedule data stream, with a planned construction period of "2024-05-10 to 2024-05-12". After segmenting the construction task name into words, the resulting word set is {concrete, pouring, main structure}. The keyword "concrete pouring" is matched in the process logic constraint rule base, with a matching degree of [missing information]. The matching degree distribution vector is {Concrete Pouring: 0.95, Rebar Tying: 0.05, Formwork Installation: 0.00}, with the maximum value corresponding to the process type "Concrete Pouring". Combined with the rule base definition that the preceding process for this process is "Rebar Tying", and the current spatiotemporal association matrix showing that rebar tying has been completed, the legality condition is met. The final output is the dominant construction process type "Concrete Pouring". This result will be combined with a high-precision positioning point set in S1.4 to construct a composite data structure, realizing the generation of semantic tags for the construction stage. In this scenario, the process determination result is completely consistent with the actual on-site conditions, ensuring the input accuracy of the dynamic threshold generation mechanism based on process semantics and significantly improving the reliability of subsequent risk assessment.

[0045] S1.4: Using the current dominant construction process type and the geometric center coordinates of the high-precision spatial positioning point set, construct a composite data structure containing process semantic descriptors and three-dimensional spatial anchor points, and convert the composite data structure into construction stage semantic tags carrying spatiotemporal attributes through an encoding mapping mechanism.

[0046] Based on the determined current dominant construction process type and the geometric center coordinates of the high-precision spatial positioning point set, the process semantic parser is invoked to encode the process type, generating a process semantic descriptor containing multi-dimensional information such as process logical attributes, work environment characteristics, and material properties. The high-precision spatial positioning point set is then input into the geometric center calculation module, and the centroid coordinates of all positioning points are calculated based on the three-dimensional Euclidean distance. The formula for the centroid coordinates is... in , , These are the three-dimensional coordinate components of the i-th positioning point. The number of positioning points is determined. The process semantic descriptor and centroid coordinates are packaged using a composite data builder to generate a composite data structure containing process semantic fields, 3D spatial anchor point fields, and timestamp fields. A hashing and positional rearrangement operation is performed on this composite data structure using an encoding mapping mechanism to form construction stage semantic tag codes that meet the standards of the stage-perturbation knowledge graph retrieval interface. Geographic reference system identifiers and construction stage serial numbers are added to the encoded semantic tags to generate construction stage semantic tags carrying spatiotemporal attributes. Through the above processing method, the process type determination and spatial positioning results of the previous step are transformed into standardized semantic tags that can directly drive knowledge graph retrieval, achieving an organic integration of construction stage identification and spatial anchoring.

[0047] For example, in a high-rise building project, a set of 450 high-precision spatial positioning points was detected during the formwork installation phase. Using the centroid calculation formula, the results were obtained... The output centroid coordinates are (54.0, 41.1, 26.7). The semantic descriptor encoding of the process includes attributes such as "formwork installation", "mixed reinforcement condition", and "lifting equipment". The composite data structure integrates the centroid coordinates and semantic descriptors and adds the current timestamp 2024-06-18 14:35:20. After transformation by the mapping mechanism, the construction stage semantic label "STG-formwork installation-54.0-41.1-26.7-T20240618143520" is obtained. This label matches the dominant disturbance source (tower crane lifting vibration, temporary load change) in the subsequent stage disturbance knowledge graph retrieval. Verification shows that the label generation is accurate and can significantly improve the stage matching degree and reduce false alarms caused by environmental disturbances.

[0048] S1.5: Perform integrity verification and format standardization on the generated construction stage semantic tags carrying spatiotemporal attributes, remove abnormal jump tags and fill in the missing timestamp field, and finally output a standardized construction stage semantic tag stream to drive the stage disturbance knowledge graph retrieval, and construct a stage disturbance knowledge graph containing working condition root nodes, disturbance source type nodes and physical constraint attribute nodes.

[0049] The stage disturbance knowledge graph is used to store structured association knowledge between different construction stages and various disturbance sources, and supports dynamic updating of association weights based on feedback data, providing S2 with the dominant disturbance source types and physical constraint attributes under the current working conditions.

[0050] This knowledge graph adopts a directed weighted graph structure and contains three types of core nodes: The root node of the work case is used to represent a specific construction stage, carrying attributes such as stage name, component type, spatial anchor point, and time window, such as "concrete pouring" and "formwork installation".

[0051] Disturbance source type nodes are used to characterize various disturbance factors that may affect structural stability, such as "pump truck vibration", "sudden load change", and "wind load".

[0052] Physical constraint attribute nodes are attached to disturbance source nodes and are used to store the quantized attributes of the disturbance source at a specific stage, such as the main frequency band range, load change amplitude, and corrosion acceleration coefficient.

[0053] The associated edges between nodes carry weight values, representing the degree of dominance or triggering probability of the disturbance source on structural stability at a specific construction stage.

[0054] The initial construction of the stage disturbance knowledge graph is based on the expert experience base, historical engineering records and industry standards, and initializes the association edges between the stage and the disturbance source and their weights.

[0055] Step S2: Based on the semantic tags of the construction stages, retrieve the pre-set stage disturbance knowledge graph, extract the dominant disturbance source type and physical constraint attributes under the current working condition to generate a stage disturbance feature vector. Specifically, this includes: S2.1: Obtain the semantic label of the construction stage as input condition, perform node matching operation to locate the root node of the working condition that completely corresponds to the label in the stage perturbation knowledge graph, thereby obtaining an initial graph subgraph containing the set of associated edges.

[0056] Based on the standardized construction stage semantic tag stream output from the preceding step S1.5, the construction stage semantic tags are used as input conditions for precise matching to initiate node matching operations. Precise node matching rules are executed in the stage perturbation knowledge graph for the input conditions, including tag text consistency verification, spatiotemporal attribute complete alignment verification, and process type code value comparison, to achieve unique location of the root node of the work condition in the graph. The matching results are limited to node instances that pass the consistency verification and appear most frequently in the tag hash index table to ensure the uniqueness and accuracy of the located nodes. For the located root node of the work condition, the edge traversal interface of the graph database is used to extract all edge sets directly associated with the node, forming an initial graph subgraph topology description containing perturbation source type nodes and physical constraint attribute nodes. The node attributes, edge types, and weight information of the initial graph subgraph are output in a structured data format, providing input for subsequent weight sorting of associated edge sets and identification of core perturbation nodes. By using the above node matching and edge set extraction processing methods, the semantic label results of the construction stage in the previous step are transformed into an initial graph subgraph that can carry stage perturbation knowledge and contains complete association information, thus realizing a precise mapping from label semantics to knowledge graph topology.

[0057] For example, in the concrete pouring stage of a high-rise building, the standardized construction stage semantic tag flow contains multi-field tags such as "stage name = concrete pouring, component type = floor slab, spatial coordinate anchor point = (125.3, 247.8, 38.5), time window = 2024-06-18T10:00~2024-06-18T16:00". The node matching rules are executed to compare the hash values ​​of the tag fields with the node index fields in the knowledge graph to ensure complete consistency, and simultaneously verify that the spatiotemporal attribute alignment error is less than [a certain value]. Rice and Using a second-based threshold, the root node of the "concrete pouring" stage was successfully located. The edge traversal interface was used to extract this root node and its associated disturbance source type nodes, such as "pump truck vibration" and "load mutation" nodes, as well as their corresponding physical constraint attribute nodes, forming an initial subgraph containing 8 associated edges, with edge weights ranging from... to The structured description of the initial subgraph includes fields such as node ID, node type, edge type, and edge weight. Subsequent steps can directly use this dataset for weight sorting and disturbance source filtering. Performance verification shows that this node matching and edge set extraction process accurately reflects the knowledge association between the construction phase and disturbance sources, and the node matching and localization time is less than [previous time]. The milliseconds significantly improve the efficiency and accuracy of generating stage perturbation feature vectors.

[0058] S2.2: Perform weighted sorting on the set of associated edges in the initial graph subgraph, and filter out high-confidence connection paths based on the confidence threshold defined by the expert experience base, so as to identify and output a list of core disturbance nodes that represent the dominant disturbance source type under the current working condition.

[0059] Based on the initial graph subgraph output by S2.1, the set of associated edges is extracted and used as the input object for weight sorting.

[0060] The edge weight calculator is invoked on the set of associated edges. Based on the connection strength values ​​recorded when the stage disturbance knowledge graph was constructed, and combined with the historical trigger frequency on the path from the root node of the current construction stage semantic label to each disturbance source node, a weight evaluation benchmark matrix is ​​generated.

[0061] The above weight evaluation benchmark matrix is ​​input into the normalized weight processing module. Standard deviation scaling is performed on the weight values ​​of different dimensions to eliminate differences in measurement scales. A correlation elimination filter is used to remove redundant highly correlated paths to ensure the independence of subsequent ranking results.

[0062] A descending sorting method is performed on the processed weight matrix to generate a sorted list from high to low weight values, and a confidence score is calculated for each connection path.

[0063] The calculated confidence score is compared with the confidence threshold preset in the expert experience base. The threshold determination module is used to filter out all high-confidence connection paths that meet the conditions, and the terminal nodes of these paths are identified as candidate disturbance source node sets.

[0064] The node mutual information analyzer is invoked on the candidate disturbance source node set to evaluate its direct correlation with the structural stability index under the current operating conditions. Nodes with insufficient correlation are eliminated to form the final core disturbance node list.

[0065] By using weighted sorting and confidence filtering, the set of associated edges of the initial graph subgraph located by S2.1 is transformed into a list of core disturbance nodes representing the dominant disturbance source type under the current construction condition, thus realizing the input conditions for subsequent physical constraint attribute extraction.

[0066] For example, during the concrete pouring stage, the set of associated edges in the initial atlas subgraph includes pump truck vibration paths, temporary surcharge paths, and wet operation paths, with edge weights of 0.82, 0.76, and 0.65 respectively, historical trigger frequencies of 18, 14, and 9 respectively, and standard deviations of all path weights of 0.05. Substituting these parameters into the formula, the confidence score of the pump truck vibration path is calculated as follows: = The confidence score for the temporary stacking path is = The confidence score for wet operation paths is = The confidence threshold of the expert experience base was set to 10. Only the pump truck vibration path and the temporary slab loading path were included in the high confidence set. The mutual information analyzer further eliminated nodes with insufficient correlation. The final output list of core disturbance nodes included pump truck vibration nodes and temporary slab loading nodes. This list was used to generate the subsequent physical constraint attribute set. The verification results showed that the model using this list significantly improved the accuracy of risk identification at this stage.

[0067] S2.3: Read the attribute dictionary of each node in the core disturbance node list, call the physical constraint parser to extract multi-dimensional physical constraint attributes such as the main frequency band range, temporary load change amplitude and corrosion acceleration coefficient, so as to generate a structured physical constraint attribute set.

[0068] An ordered traversal of the attribute dictionary of each node in the core disturbance node list is performed, and the multidimensional record entry containing frequency spectrum data, load change information and corrosion parameters is located in the attribute dictionary using the node's unique identifier as the search key.

[0069] The frequency domain decomposition module of the preset physical constraint resolver is called to extract the spectral peak positions and calculate the main frequency band range for the frequency spectrum data in the attribute dictionary. The spectral peak detection method is used to identify the frequency interval where the maximum power is located and output the start and end frequency values ​​of the interval.

[0070] The load change information in the attribute dictionary is used to call the load mutation analysis module of the physical constraint resolver. The instantaneous change amplitude is calculated based on the time series load data. The mutation value is obtained through the differential operator and the maximum mutation amount is used as the temporary load mutation amplitude.

[0071] The environmental action parsing module of the physical constraint parser is called on the corrosion parameters in the attribute dictionary. The corrosion acceleration coefficient is calculated by combining historical temperature and humidity conditions. The weighted average method is used to fuse the multi-stage corrosion rates to obtain the corrosion acceleration coefficient value of the structure under the current working conditions.

[0072] The aforementioned frequency band range, temporary load change amplitude, and corrosion acceleration coefficient are sequentially encapsulated into a unified data structure, and attribute names and values ​​are labeled to form a structured physical constraint attribute set with clear field definitions and numerical domains.

[0073] By calling the standardized output interface of the physical constraint parser, the result of the previous step is transformed into a set of structured physical constraint attributes, realizing a unified expression of multi-dimensional physical constraint features for the current working condition, which facilitates subsequent vectorized encoding and feature fusion processing.

[0074] For example, in the implementation of edge protection monitoring for high-altitude operations, the attribute dictionary of the core disturbance node already contains three types of data: the first type is the vibration frequency spectrum, with a sampling frequency of 200Hz, and the main frequency range calculated by the physical constraint resolver is... Hz; the second category is load variation information, with a time series length of 120 seconds. The maximum abrupt change amplitude is obtained using the differential operator. kN; the third category is corrosion parameters, with a historical average temperature of ℃, humidity The corrosion acceleration coefficient was obtained by weighted average method. These three types of attributes, after being encapsulated in a unified data structure, form a structure with a field "freq". band = "load" spike = "corrosion" coeff = The structured physical constraint attribute set. In the subsequent vectorized encoding stage, this attribute set can be used as input to significantly improve the engineering context accuracy of feature mapping and enhance the adaptability of stage perturbation features during dynamic threshold generation.

[0075] S2.4: Based on the structured physical constraint attribute set, perform vectorized encoding operation, and use a domain-specific multidimensional feature embedding method to map discrete attributes into dense vectors in continuous numerical space to output a preliminary fused engineering context feature vector.

[0076] S2.5: Normalize and orthogonalize the initially fused engineering context feature vectors, use standard deviation scaling to eliminate dimensional differences and correct the coupling effect between features, so as to finally generate a standardized stage perturbation feature vector for retrieving historical time series curves.

[0077] like Figure 2 As shown, step S3 involves: using the stage disturbance feature vector to retrieve and match the parameter time-series curves of similar projects with the current working conditions from the historical building engineering monitoring database; and constructing a benchmark stability decay trajectory template characterizing the structural response law of this construction stage after normalization. Specifically, this includes: S3.1: Obtain the full set of time-series curves from the pre-set historical building engineering monitoring database, perform multi-dimensional spatial similarity matching operation based on the stage disturbance feature vector, calculate the Euclidean distance score between each historical time-series curve and the current working condition feature, and screen out a cluster of high-confidence candidate time-series curves with similar topological structure and consistent physical properties.

[0078] The historical building engineering monitoring database is used to store the time-series monitoring data of completed projects and the corresponding engineering topology parameters, providing S3 with historical response curves similar to the current working conditions to construct a benchmark stability decay trajectory template.

[0079] The database uses time-series curves as its core storage unit. Each curve is associated with the following metadata: project identification fields, including project name, construction stage, component type, etc.; topological parameter fields, including component connection diagram (node-edge relationship), support system type, boundary conditions, etc.; physical constraint attribute fields, including frequency band range, load change amplitude, corrosion acceleration coefficient, temperature and humidity coupling strength, etc.; time-series data fields, which store the structural response amplitude sequence along the time axis, covering physical quantities such as displacement, stress, and vibration; and sampling information fields, including sampling frequency, time window length, and sensor type of data source, etc.

[0080] The historical building engineering monitoring database collects raw data through the on-site monitoring system of historical engineering projects. After cleaning, normalization, and time axis alignment, the data is stored in the database. When each curve is stored, its topological parameters and physical constraint attributes are automatically extracted to establish a multi-dimensional index to support S3.1 similarity matching.

[0081] Based on the standardized stage disturbance feature vector output from the previous steps, this feature vector is used as a query condition input into a pre-set historical building engineering monitoring database to limit the access scope of the full set of time-series curves. Multidimensional feature mapping is performed on the metadata of each time-series curve in the database, strictly aligning the engineering topology parameters corresponding to the curve with the physical constraint attribute dimension of the stage disturbance feature vector to ensure the comparability and dimensional consistency of the comparison parameters. The Euclidean distance calculation method is used to evaluate the similarity between historical time-series curves and current operating condition characteristics line by line in a multidimensional continuous numerical space, and the distance score is calculated using the following formula: in, This represents the attribute value of the i-th historical curve in the j-th dimension. This represents the attribute value of the current operating condition feature vector in the j-th dimension, where m is the total number of feature dimensions. The calculated Euclidean distance scores are sorted from smallest to largest, and curves with distance scores below a preset similarity threshold are selected as high-confidence candidate curve clusters. Topological consistency verification is performed on the selected curve clusters by matching the component connection maps of each curve with the current stage's structural model, eliminating curve samples with inconsistent structural connection patterns. By combining multidimensional spatial similarity matching and topological consistency verification, the stage perturbation feature vector from the previous step is transformed into a high-confidence candidate time-series curve cluster with similar topological structures and consistent physical properties, providing the data foundation required for the subsequent construction of the baseline stability decay trajectory template.

[0082] For example, in a high-altitude edge protection monitoring scenario, the standardized disturbance feature vector m is set to 6 dimensions, including the average frequency band (14.2Hz), temporary load abrupt change amplitude (4.6kN), corrosion acceleration coefficient (0.025), peak displacement offset (3.1mm), temperature and humidity coupling strength (0.62), and bolt preload attenuation rate (0.012). A total of 120 time-series curves are extracted from the historical building engineering monitoring database. The Euclidean distance between each curve and the current working condition characteristics in the above 6 dimensions is calculated, with the threshold set as follows: A total of 42 curves with distance scores below this value were selected. Consistency verification was performed on these 42 candidate curves regarding their component connection patterns and the current stage structural model. Fifteen curves with significantly different connection patterns were removed, resulting in 27 high-confidence candidate curves. This cluster of candidate curves was used for time-axis resampling and normalization in subsequent steps. Verification results showed that the trajectory template constructed based on this cluster exhibited a significantly reduced fitting error between the stability decay trend in the prediction stage and the measured curves, achieving high-precision extraction of stage response patterns.

[0083] S3.2: Perform time axis resampling processing on each original time series curve in the high confidence candidate time series curve cluster, map historical data with different sampling frequencies to a standard time grid, and perform dimensionless scaling on the amplitude domain based on the maximum and minimum value normalization method to eliminate dimensional differences and generate a standardized historical response sequence set.

[0084] S3.3: Based on the standardized historical response sequence set, perform dynamic time warping and alignment operation to identify the characteristic peak moments of key turning points representing structural stability in each sequence, and correct time offset errors through linear interpolation to strictly synchronize all standardized historical response sequences to a unified evolution phase reference in the time dimension, generating a family of phase-aligned time series curves.

[0085] S3.4: Perform point-by-point statistical aggregation operation on the phase-aligned time series curve family, calculate the arithmetic mean of the amplitude of all curves at each time slice point as the central trend term, and calculate the standard deviation as the discrete fluctuation term to quantify the general law and random fluctuation range of the structural response under this construction stage, and generate a preliminary statistical template containing the mean trajectory line and the variance envelope.

[0086] Perform point-by-point aggregation statistics on the amplitude sequence of each curve in the phase-aligned time series curve family, and arrange the amplitudes of multiple curves at the same time slice point into a matrix row vector according to the index consistency principle.

[0087] The arithmetic mean of the column directions of the row vectors of the above matrix is ​​used to obtain the sequence of central trend terms.

[0088] Perform standard deviation calculation on the amplitude set at the same time slice point to obtain the discrete fluctuation term sequence.

[0089] The central trend term sequence is arranged in chronological order to form the mean trajectory line, and the discrete fluctuation term sequence is used as the upper and lower envelope amplitudes to construct the variance envelope band.

[0090] The mean trajectory and variance envelope are associated and stored to form a preliminary statistical template containing the mean trajectory and variance envelope hyperbola.

[0091] Through the above statistical aggregation processing method, the amplitude data after phase alignment in the previous step is transformed into an index that quantifies the general laws and random fluctuation range of structural response, thereby realizing a visual representation of the stability evolution characteristics during the construction stage.

[0092] For example, during the concrete pouring stage, the cluster of high-confidence candidate time-series curves is configured with 10 curves, with a uniform resampling frequency of 1Hz and a standard time grid length of 1200 seconds. In the aggregation operation, each time slice contains 10 amplitude samples, and the arithmetic mean formula... =10. By averaging point by point, the amplitude sequence of the mean trajectory line shows a trend of gradually decreasing from 0.05m to 0.01m within 0~1200 seconds. The standard deviation formula shows a fluctuation range of 0.015m in the initial stage of pouring (0~300 seconds), decreasing to 0.005m in the middle stage (300~900 seconds), and stabilizing at 0.002m in the final stage. The mean trajectory line and the standard deviation double envelope are plotted to form a preliminary statistical template, which can be used for subsequent cubic spline interpolation smoothing. Verification shows that the template can significantly improve the accuracy of stability risk interval determination and reduce threshold misjudgment caused by environmental noise in comparison with real field data.

[0093] S3.5: The mean trajectory line in the preliminary statistical template is smoothed and fitted using the cubic spline interpolation method to remove local high-frequency jitter caused by historical data noise. The variance envelope is then expanded and corrected using a preset confidence coefficient, and finally a smooth, continuous, and statistically significant baseline stability decay trajectory template is output.

[0094] The input condition for the mean trajectory line in the preliminary statistical template after phase alignment is time series data generated by S3.4, which includes the arithmetic mean of amplitude and the variance envelope.

[0095] Based on the discrete sampling points of the mean trajectory, cubic spline interpolation is performed to construct the original discrete points into a piecewise cubic polynomial function. Each polynomial is continuous between adjacent sampling points and its first and second derivatives are continuous, ensuring the smoothness of the curve.

[0096] In each interpolation segment, the first derivative value at the endpoint is defined using boundary conditions to constrain the curve to prevent non-physical spikes from appearing globally, thereby eliminating local high-frequency jitter caused by historical data noise.

[0097] By combining the variance envelope in the preliminary statistical template, the standard deviation of each time slice is read and used as a quantitative indicator of random fluctuation amplitude. A pre-set confidence coefficient is introduced to expand and correct the variance envelope, ensuring that the template covers all high-confidence historical evolution trajectories.

[0098] The variance envelope expansion factor is calculated. Based on the upper and lower bounds of the expanded envelope, the mean trajectory line after smooth interpolation is embedded into it to generate a smooth, continuous, and statistically significant baseline stability decay trajectory template.

[0099] By using cubic spline interpolation and variance confidence expansion, the statistical template from the previous step is transformed into continuous function data containing mean trend and high-confidence fluctuation range, thus realizing a benchmark stability decay trajectory template that can be directly called by the subsequent dynamic threshold generation mechanism.

[0100] For example, in a certain rebar binding stage, the input conditions are the mean trajectory points and corresponding standard deviations of 30 time slices after phase alignment. The amplitude of the mean trajectory points ranges from 0.85 to 0.92, and the standard deviation ranges from 0.004 to 0.007. When performing cubic spline interpolation, the polynomial coefficients for each segment are obtained by solving the tridiagonal matrix, with the first derivative constraint at the endpoints set to zero, and the interpolation curve has no sharp peaks across the entire domain. A pre-set reliability coefficient is also provided. At a 95% confidence level, the standard deviation at the 15th time slice is 0.006, therefore the expansion factor is... The upper bound extension value is obtained as 0.01176. The mean trajectory line is combined with the upper and lower edge envelopes to form a smooth and continuous template. In practical applications, this template can effectively improve the stability of risk assessment in the binding stage and significantly improve the accuracy of early warning judgment when generating the subsequent dynamic threshold envelope.

[0101] like Figure 3 As shown, step S4 involves: acquiring real-time sensor data streams and calculating vibration energy entropy, displacement trend slope, and temperature and humidity coupling offset, which are then input into a long short-term memory network as short-term disturbance input conditions to output residual compensation coefficients. Specifically, this includes: Real-time sensor data streams are acquired through a multi-type sensor network pre-deployed on the edge protection structure. The sensor network includes accelerometers, displacement sensors, temperature sensors, and humidity sensors, and the arrangement of each sensor is as follows: Accelerometers are positioned at the top of the edge protection uprights and at the mid-span of the crossbars to collect vibration response signals of the structure under construction disturbances. Each monitoring section is equipped with no fewer than two accelerometers, with a sampling frequency of at least 1024Hz and a measurement range set to ±2g to ±10g based on the estimated vibration amplitude. The accelerometers are secured using magnetic or bolt methods to ensure a rigid connection to the structure and to obtain accurate vibration transmission characteristics.

[0102] Displacement sensors, employing either wire-type or laser displacement sensors, are deployed at deformation-sensitive locations on the edge protection structure, including the connection nodes between uprights and floor slabs, mid-span of horizontal members, and corners. These sensors monitor the relative displacement changes of the structure under load, with a sampling frequency of at least 100Hz. The measurement range is set according to the allowable deformation value in the structural design, and the measurement accuracy is better than ±0.1mm. During installation, the displacement sensor's measurement direction must be aligned with the main deformation direction of the structure to avoid measurement errors introduced by installation angle deviations.

[0103] Temperature and humidity sensors are arranged in pairs near the edge protection area, installed approximately 1.5 meters above the work surface to avoid direct sunlight and interference from localized heat sources. The temperature sensor measures from -20℃ to 60℃ with an accuracy better than ±0.5℃; the humidity sensor measures from 0% to 100%RH with an accuracy better than ±3%RH. The temperature and humidity sensors have a sampling period of 1 second, and the data is synchronized with the structural response signal to calculate the temperature-humidity coupling strength and corrosion acceleration factor.

[0104] All sensors are connected to the field data acquisition terminal via wired or wireless means. The acquisition terminal has multi-channel synchronous sampling capabilities and supports timestamping of data from each sensor according to a unified time base. The data acquisition terminal transmits the raw sensor data stream to the edge computing unit in real time for feature extraction and residual compensation coefficient calculation.

[0105] S4.1: Acquire the original timing signals of the accelerometer and displacement sensor, perform sliding window segmentation on the original timing signals of the accelerometer to generate segmented acceleration data blocks, and perform fast Fourier transform operation on the segmented acceleration data blocks to extract the frequency domain power spectral density distribution. Then, use the Shannon entropy calculation formula to perform disorder quantification on the frequency domain power spectral density distribution to obtain the vibration energy entropy value.

[0106] Long Short-Term Memory (LSTM) networks are used to learn the nonlinear mapping relationship between short-term environmental disturbances (vibration energy entropy, displacement trend slope, and temperature and humidity coupling offset) and structural stability residuals, outputting residual compensation coefficients for use in S5 dynamic threshold fusion. This network employs a lightweight design to reduce computational overhead and adapt to edge computing environments at construction sites.

[0107] The input layer receives a three-dimensional tensor (T,F,B), where T is the time step, F=3 is the number of features (vibration energy entropy, displacement trend slope, temperature and humidity coupling offset), and B is the batch size.

[0108] The LSTM layer adopts a single-layer structure with H hidden units (typically 32 to 64), including forget gate, input gate, and output gate, and the number of parameters is approximately 4×(H×(F+H)+H).

[0109] The fully connected layer maps the hidden state to a scalar output with H+1 parameters.

[0110] The output layer output residual compensation coefficient is a scalar value.

[0111] This network employs a single-layer LSTM structure with a small number of hidden units, keeping the total number of parameters in the thousands. The input feature dimension is fixed at 3, while the time step T can be flexibly configured according to the working conditions (e.g., 10 to 20). The network supports real-time inference on embedded edge devices, meeting the low-latency response requirements of construction sites.

[0112] The original time-series signals from the accelerometer and displacement sensor are acquired synchronously to ensure the consistency of the sampling time reference for subsequent joint analysis.

[0113] The raw time-series signal from the accelerometer is processed by sliding window segmentation, which divides the continuous sampling sequence into several segmented acceleration data blocks according to a fixed window length and step size, in order to capture local dynamic features and avoid diluting short-term disturbance characteristics by long-span time-series information.

[0114] A fast Fourier transform is performed on each segment of the acceleration data block to obtain the corresponding frequency domain power spectral density distribution. The frequency domain transformation is used to analyze the energy distribution characteristics of the signal at different frequency components.

[0115] The vibration energy entropy value is calculated based on the frequency domain power spectral density distribution using the Shannon entropy formula: in, The entropy value represents the energy proportion of the i-th frequency point in the normalized power spectral density. The entropy value characterizes the degree of disorder in the frequency domain energy distribution of the vibration signal. The larger the value, the more uniform the frequency domain energy distribution and the more obvious the information entropy characteristics of short-term disturbances.

[0116] The calculated vibration energy entropy value is used as part of the short-term perturbation feature vector for subsequent input to the lightweight long short-term memory network.

[0117] By using sliding window segmentation, fast Fourier transform, and Shannon entropy calculation, the original acceleration time series signal is transformed into vibration energy entropy data required for structural stability risk assessment, thereby achieving a quantitative characterization of the frequency domain disorder of short-term disturbances.

[0118] For example, in a high-altitude edge protection monitoring scenario, the accelerometer sampling frequency is configured to 1024Hz, the sliding window length is set to 2048 points, and the step size is set to 256 points. A fast Fourier transform is performed on each window to obtain the power spectral density distribution of 1024 frequency points. After normalization, the energy percentage of the top 100 frequency points is calculated. The vector is then used, and the Shannon entropy formula is applied. Since the energy concentration in the 12Hz~18Hz main frequency band reaches 0.6, while the energy concentration in other frequency bands is relatively low, the entropy value is calculated as follows: This indicates that the vibration energy distribution at this stage exhibits a dominant concentration in the frequency domain. Furthermore, comparing the entropy values ​​with other stages reveals a significant decrease in entropy during the concrete pouring stage, demonstrating the influence of different processes on frequency domain disorder. This entropy value can provide the LSTM model with key input parameters reflecting the current short-term disturbance characteristics, significantly improving the stage adaptability of dynamic threshold generation.

[0119] S4.2: Obtain the original time-series signals of the displacement sensor within the same time window, perform first-order difference operation on the original time-series signals of the displacement sensor to generate a displacement change rate sequence, and perform least squares linear fitting processing based on the displacement change rate sequence to extract the linear regression slope parameter, thereby determining the linear regression slope parameter as the displacement trend slope characterizing the structural deformation trend.

[0120] S4.3: Acquire ambient temperature data collected by the temperature sensor and ambient relative humidity data collected by the humidity sensor. Perform normalization preprocessing on the ambient temperature data and ambient relative humidity data to generate standardized temperature and humidity variables. Calculate the Pearson correlation coefficient based on the standardized temperature and humidity variables to quantify the temperature and humidity coupling strength. Then, combine the preset corrosion acceleration factor mapping table to perform weighted fusion processing on the temperature and humidity coupling strength to obtain the temperature and humidity coupling offset.

[0121] The corrosion acceleration factor mapping table is used to establish a quantitative correspondence between "temperature and humidity coupling strength" and "structural corrosion acceleration effect". Since the combined effect of temperature and humidity has a nonlinear accelerating effect on the corrosion rate of metal components, the temperature and humidity coupling strength coefficient alone cannot directly characterize its influence on structural stability. Therefore, this mapping table is needed to convert the temperature and humidity coupling strength into a corrosion acceleration factor that can participate in subsequent fusion calculations.

[0122] This mapping table is built based on the following three types of data sources: Electrochemical theoretical model: Based on the electrochemical principle of metal corrosion, the corrosion rate and temperature and relative humidity satisfy the Arrhenius type relationship. In the range of relative humidity of 60% to 90%, the corrosion rate increases exponentially with the increase of humidity.

[0123] Accelerated aging test data: Through laboratory accelerated corrosion tests (salt spray test, temperature and humidity cycle test), the corrosion depth growth rate of standard specimens under different temperature and humidity combinations was measured, and an empirical relationship between temperature and humidity coupling strength and acceleration factor was established.

[0124] Historical engineering measured data: Corrosion detection data of similar components in completed projects and environmental temperature and humidity records during the same period were collected. The statistical mapping relationship between temperature and humidity coupling strength and corrosion acceleration coefficient was obtained through regression analysis.

[0125] The mapping table uses discrete intervals of the temperature and humidity coupling strength coefficient ρ as indexes. Each interval corresponds to a corrosion acceleration factor k, which is a dimensionless multiple, representing the amplification factor of the corrosion rate under the temperature and humidity conditions relative to the standard operating conditions.

[0126] The specific mapping relationship is as follows: when the temperature and humidity coupling strength ρ is less than 0.2, the corrosion acceleration factor k is 1.0; when ρ is in the range of 0.2 to 0.4, k is 1.2; when ρ is in the range of 0.4 to 0.6, k is 1.5; when ρ is in the range of 0.6 to 0.8, k is 1.8; when ρ is greater than or equal to 0.8, k is 2.2.

[0127] Among them, the temperature and humidity coupling strength ρ is the absolute value of the Pearson correlation coefficient r, which ranges from [0,1]. The larger the value, the more significant the synergistic change in temperature and humidity, and the stronger the accelerating effect on corrosion.

[0128] The corrosion acceleration factor k obtained from the lookup table is weighted and fused with the temperature and humidity coupling strength ρ to generate the final temperature and humidity coupling offset.

[0129] This mapping table is not fixed and can be periodically revised based on on-site handling records and corrosion detection data fed back by S7. For example, if multiple warnings within a certain temperature and humidity coupling intensity range are confirmed to be false alarms (actual corrosion did not meet expectations), the acceleration factor k for that range is reduced accordingly; conversely, if a false alarm occurs (actual corrosion exceeded expectations but no warning was issued), k is increased. Through this mechanism, the mapping table achieves adaptive optimization to the engineering environment.

[0130] The ambient temperature data collected by the temperature sensor and the ambient relative humidity data collected by the humidity sensor are clock-synchronized to ensure the correspondence of data within the same time window.

[0131] Linear normalization is performed on the ambient temperature data, mapping the original temperature values ​​to a dimensionless standardized range of [0,1] according to a preset physical quantity range, thus forming a standardized temperature variable.

[0132] The same linear normalization process is performed on the relative humidity data to map the humidity percentage to the [0,1] interval, forming a standardized humidity variable.

[0133] A bivariate sample sequence was constructed based on standardized temperature and humidity variables, and the coupling strength between the two variables was quantified using the Pearson correlation coefficient formula; the coefficient calculation formula is as follows: in To standardize temperature values, To standardize humidity values, and These are the average values ​​for temperature and humidity, respectively.

[0134] The obtained temperature and humidity coupling strength coefficient is used as the input index to find the corresponding acceleration factor in the preset corrosion acceleration factor mapping table, and then weighted fusion processing is performed.

[0135] Through the above processing method, the raw data of the temperature and humidity sensor is transformed into a temperature and humidity coupling offset that reflects the intensity of the interaction between environmental temperature and humidity and combines the effect of structural corrosion, so as to realize the quantitative input of short-term environmental disturbances.

[0136] For example, in a high-altitude edge protection monitoring scenario, the temperature sensor collects data ranging from 21.5℃ to 28.7℃, and the humidity sensor collects data ranging from 68% to 85%. The sampling period is 1 second, and the temperature and humidity data are synchronized with a clock. The temperature normalization interval is set to [20,30]℃ mapped to [0,1], and the humidity normalization interval is set to [60,90]% mapped to [0,1]. After normalization, the mean value of the temperature variable is 0.45, and the mean value of the humidity variable is 0.72. The results are calculated using the Pearson correlation coefficient. = In the corrosion acceleration factor mapping table, this coefficient corresponds to the factor value. = The temperature and humidity coupling offset is obtained by performing weighted fusion. = This value, as an input feature, can significantly improve the sensitivity of lightweight long short-term memory networks to short-term environmental disturbances, thereby optimizing the generation accuracy of personalized dynamic thresholds and significantly reducing the risk of misjudgment caused by temperature and humidity fluctuations during validation.

[0137] S4.4: Obtain the vibration energy entropy, displacement trend slope, and temperature and humidity coupling offset generated in the previous steps. Align and stitch the vibration energy entropy, displacement trend slope, and temperature and humidity coupling offset according to the timestamp to construct a multidimensional short-term disturbance feature vector. Perform tensor reshaping operation based on the multidimensional short-term disturbance feature vector to generate three-dimensional tensor data that meets the input requirements of lightweight long short-term memory networks.

[0138] The vibration energy entropy, displacement trend slope, and temperature and humidity coupling offset output from the previous steps are used as the original input conditions for the multi-source short-term disturbance parameters. The timestamp index of each parameter is defined and a unified time axis reference system is established to ensure the accurate alignment of multi-dimensional features in temporal relationships.

[0139] The vibration energy entropy, displacement trend slope, and temperature and humidity coupling offset are subjected to timestamp index matching operation to map the feature values ​​at different sampling frequencies to a unified sampling node, thereby eliminating the time offset caused by the difference in sampling period.

[0140] Dimensional concatenation is performed based on the aligned feature value sequence. Perturbation residual feature vectors containing three components are formed at each time slice node according to the predefined feature arrangement order, ensuring that the vector structure is fixed and consistent with the input dimension of the subsequent network.

[0141] Tensor dimension expansion is performed on the multidimensional perturbation residual feature vector, with the time axis as the first dimension, the feature components as the second dimension, and a single batch flag as the third dimension, initially constructing a structure of the form ( The three-dimensional tensor skeleton of ) For the number of time steps, The number of characteristic components, This refers to the batch size.

[0142] The numerical values ​​in the 3D tensor skeleton are standardized and scaled, and mean-variance normalization is used to eliminate the interference of different units and numerical scales on the stability of network training.

[0143] Through the above alignment, splicing, expansion and normalization processing methods, the multi-source perturbation parameters of the previous step are transformed into three-dimensional tensor data that meet the input requirements of lightweight long short-term memory networks and have time series structure characteristics, so as to realize a unified formatted expression of short-term perturbations.

[0144] For example, in a monitoring scenario during the concrete pouring stage, the vibration energy entropy calculation period is 200ms, the displacement trend slope calculation period is 500ms, the temperature and humidity coupling offset update period is 1s, and the sampling timestamps are 0.2, 0.4, and 0.6 seconds, respectively. Through timestamp matching operations, the displacement trend slope and temperature and humidity coupling offset are interpolated to the vibration energy entropy sampling node, forming a disturbance residual feature vector containing three components at each time point. For example, at 0.4 seconds, the corresponding vector is [1.32, [0.015, 0.08]. When constructing the 3D tensor, the time step is set to... =10, eigencomponents =3, batch size =1, after normalization, the mean of the vibration energy entropy component is 1.25, the standard deviation is 0.08, and the normalized value at 0.4 seconds is... =0.875. The tensor structure fully meets the input requirements of lightweight long short-term memory networks, and the network can significantly improve the accuracy of residual compensation coefficient prediction and the ability to adapt to environmental disturbances during the inference stage.

[0145] S4.5: Obtain the preset long short-term memory network model parameters, input the three-dimensional tensor data into the input gate layer of the long short-term memory network model for forgetting gating update processing to generate cell state update quantity, and perform output gate activation function mapping operation based on the cell state update quantity to extract hidden layer state features. Finally, perform linear regression prediction processing on the hidden layer state features through a fully connected layer to output scalar form residual compensation coefficient.

[0146] The three-dimensional tensor data output from the previous steps is used as the input object of the lightweight long short-term memory network (LSTM) model, and the preset network weights and structural configuration parameters are called to initialize the internal computation state.

[0147] In the input gate layer, matrix multiplication mapping is performed on the three-dimensional tensor data. Candidate state vectors are generated by combining the model's preset input weight matrix and bias vector. The Sigmoid function is then used to normalize and compress the mapping output to form the input gate activation value.

[0148] Based on the product of the input gate activation value and the cell state at the previous time step, combined with the weight matrix and activation value of the forget gate, the forgetting and retention of the cell state is controlled by element-wise multiplication, and the cell state update amount is calculated.

[0149] In the output gate calculation, a preset output weight matrix is ​​combined and mapped with the current input data and cell state update, and the output gate activation value is obtained after Sigmoid activation. The hyperbolic tangent function is then applied to the cell state update to obtain the nonlinear change amplitude of the cell state.

[0150] The hidden layer state features are input into the fully connected layer to perform linear regression prediction, and the hidden state is weighted and summed using the weight vector and bias term.

[0151] Through the aforementioned gating mechanism and fully connected mapping processing, the multidimensional short-term disturbance feature tensor of the previous step is transformed into a scalar form of residual compensation coefficient, thereby realizing a quantitative characterization of the impact of short-term environmental disturbances on structural stability during the current construction phase.

[0152] For example, during the concrete pouring stage of a high-rise building, the three-dimensional tensor data includes vibration energy entropy of 0.85, displacement trend slope of 0.012, and temperature and humidity coupling offset of 0.15, corresponding to a time series length of 50. The input gate weight matrix is ​​(32×3) in size, the bias vector length is 32, the forget gate weight matrix is ​​(32×3) with a bias length of 32, and the output gate weight matrix is ​​(32×3) with a bias length of 32. After input gate sigmoid normalization, the input tensor yields an average activation value of approximately 0.74, and the forget gate activation value is approximately 0.68. Combined with the cell state of 0.21 from the previous time step, the cell state update is calculated to be approximately 0.36. The output gate activation value is approximately 0.71, and the hyperbolic tangent mapping value of the cell state is approximately 0.345. The element-wise multiplication of these two values ​​yields the hidden layer state feature of approximately 0.245. The weight matrix of the fully connected layer is (1×32) in size and has a bias value of 0.005. After weighted summation of the hidden layer state feature vectors, the residual compensation coefficient is approximately 0.052. When applied to dynamic threshold fusion, it significantly improves the sensitivity and adaptability to short-term environmental disturbances.

[0153] Step S5: The predicted value of the baseline stability decay trajectory template is nonlinearly fused with the residual compensation coefficient to generate a dynamic threshold envelope for the current construction stage. Specifically, this includes: S5.1: Based on the output node of the baseline stability decay trajectory template at the current moment, extract the baseline stability prediction amplitude that characterizes the natural evolution trend of the structure, and perform dimensional normalization processing on the baseline stability prediction amplitude to eliminate the interference of different physical magnitudes on subsequent fusion operations, and generate a standardized baseline stability prediction vector.

[0154] Using the output node of the baseline stability decay trajectory template at the current moment as the starting point for calculation, the predicted amplitude data corresponding to that node is read to form the original amplitude sequence characterizing the natural evolution trend of the structure.

[0155] The original amplitude sequence is subjected to unit physical quantity detection to identify the dimensional category of the data item, including meters, millimeters, Newtons, or megapascals, in order to establish a dimensional feature mapping table.

[0156] Based on the dimensional feature mapping table, the amplitude data is transformed to a preset reference dimension to eliminate the influence of physical unit differences on numerical comparison.

[0157] The converted amplitude sequence is processed using a normalization method, while strictly adhering to the grouping structure.

[0158] During the normalization process, the arithmetic mean and standard deviation of the amplitude sequence are calculated, and each term is converted into a dimensionless standardized predicted value according to the formula.

[0159] The standardized set of predicted values ​​is constructed into a standardized baseline stability prediction vector, providing a unified scale input for subsequent fusion with the residual compensation coefficients.

[0160] Through the above processing method, the original predicted amplitude of the baseline stability decay trajectory template is transformed into a standardized baseline stability prediction vector with dimensionless uniformity that can directly participate in multidimensional nonlinear fusion calculations, thereby achieving fusion and adaptation of data of different physical magnitudes.

[0161] For example, in the monitoring of edge protection during high-altitude operations, the baseline stability decay trajectory template outputs node prediction amplitudes of 0.015 meters, 12.5 Newtons, and 0.003 MPa at the current moment, representing displacement, bolt preload, and component stress, respectively. An identity conversion is performed from meters to meters for displacement data, from Newtons to MPa for bolt preload (with an area of ​​0.001 square meters), and from MPa to MPa for component stress. The converted values ​​are unified as follows: meters = 0.015 meters, Newtons = 12.5 / 1000 = 0.0125 MPa, and MPa = 0.003 MPa. After unifying all dimensions, the average value of the unified sequence is calculated. =0.0105, standard deviation =0.0053. Using normalization, the normalized values ​​for each term are calculated as follows: displacement term normalized value = (0.015 - 0.0105) / 0.0053 ≈ 0.849, preload term normalized value = (0.0125 - 0.0105) / 0.0053 ≈ 0.377, stress term normalized value = (0.003 - 0.0105) / 0.0053 ≈ -1.415. This standardized baseline stability prediction vector [0.849, 0.377, -1.415] significantly improves the fusion accuracy between different physical quantity indicators and the reliability of risk assessment when subsequently fused with the residual compensation coefficient.

[0162] S5.2: Using the residual compensation coefficients output by the Long Short-Term Memory Network, combined with a preset family of nonlinear activation functions, perform dynamic weight mapping processing on the residual compensation coefficients to quantify the instantaneous impact intensity of short-term environmental disturbances on structural stability and generate a disturbance sensitivity weighting factor.

[0163] Using the residual compensation coefficients output from the preceding steps as input, the system calls multiple candidate function libraries from the pre-set nonlinear activation function family. Within the function selection module, the absolute value and sign of the residual compensation coefficients are used as dual-condition triggering parameters to determine the applicable function type and corresponding parameterization template.

[0164] The residual compensation coefficient is input into the selected nonlinear activation function, and a point-by-point mapping operation is performed to obtain the disturbance response value within the unit normalization range. The amplitude weight of the stage disturbance feature vector is superimposed on the mapping result to realize the first-order coupling between the intensity of environmental disturbance and the sensitivity of structural stability.

[0165] A dynamic weight adjustment factor is introduced into the coupling operation results. This adjustment factor is calculated based on the variance ratio of the real-time environmental parameter volatility to the historical disturbance residual distribution.

[0166] The dynamic weight adjustment factor is applied to the perturbation response value, and a weighted perturbation sensitivity vector is generated through dot product operation. A range pruning operation is performed on the sensitivity vector to limit the value within the nonlinear stable range, so as to prevent abnormal drift of the upper or lower bound of the threshold caused by extreme input.

[0167] A smoothing filter is applied to the cropped sensitivity vector to perform a first-order low-pass filter on the sensitivity changes in continuous sampling periods, suppressing high-frequency noise components and ensuring the stability of subsequent threshold fusion operations.

[0168] Through the above dynamic weight mapping and smoothing constraint processing, the residual compensation coefficient of the previous step is transformed into a disturbance sensitivity weighting factor that can measure the instantaneous impact intensity of short-term environmental disturbances, thereby achieving a balanced output of construction environment adaptability and risk sensitivity.

[0169] For example, during the concrete pouring stage of a high-rise building, the output residual compensation coefficient of the lightweight long short-term memory network is 0.35. The function selection module selects the hyperbolic tangent function template based on the positive value of the coefficient, with parameters set to a gain coefficient of 1.2 and an offset of 0. The perturbation response value after mapping is calculated as follows: The result is 0.4027. The standard deviation of the real-time environmental parameter volatility in the current stage is 0.15, and the standard deviation of the historical disturbance residuals in the same stage is 0.10. The dynamic weight adjustment factor is calculated as follows: =1.5. The weighted disturbance sensitivity vector is 0.4027×1.5=0.60405. After range clipping and constraint to the interval [-0.8,0.8], we get 0.60405. The low-pass filter is set to a cutoff frequency of 0.25Hz. The sensitivity data of three consecutive cycles are smoothed to obtain the final disturbance sensitivity weighting factor of 0.588. This factor shows a significant improvement in the environmental disturbance response capability in the subsequent coupling gain matrix generation, ensuring the dynamic matching of the threshold envelope with the actual risk state.

[0170] S5.3: Based on the corresponding dimensional relationship between the perturbation sensitivity weighting factor and the standardized baseline stability prediction vector, construct a coupling gain matrix and perform eigenvalue decomposition on the coupling gain matrix to identify the key modal components that dominate the stability risk in the current stage and generate modal decoupling correction parameters.

[0171] The perturbation sensitivity weighting factor is matched with the standardized baseline stability prediction vector in the corresponding dimension to form the element mapping rule of the input matrix.

[0172] According to the matching rules, the perturbation sensitivity of each dimension is multiplied with the corresponding baseline stability prediction value to generate an initial gain matrix containing the product results of the weights of each dimension and retaining the dense structure of the matrix.

[0173] An adaptive adjustment mechanism is applied to the initial gain matrix, which performs nonlinear scaling on the matrix elements according to preset sensitivity amplification and suppression coefficients to form a coupled gain matrix that conforms to the risk characteristics of the current construction stage.

[0174] Based on the coupling gain matrix, its covariance matrix is ​​constructed, and the corresponding eigenvalues ​​and eigenvectors are solved using the eigenvalue decomposition method to identify the principal and minor components of the matrix.

[0175] Using the principal component screening criterion, the parts of all eigenvalues ​​greater than the threshold are selected as the key modal components that dominate the current stability risk. Modal decoupling correction parameters are extracted based on the corresponding eigenvectors for subsequent nonlinear fusion calculations.

[0176] By using the matrix construction and eigenvalue decomposition methods described above, the perturbation sensitivity and the baseline prediction vector are transformed into a set of modified parameters that characterize the risk-dominant mode, thereby achieving precise tuning of the stability assessment criteria under different perturbation effects.

[0177] For example, in a high-altitude edge protection monitoring scenario, the disturbance sensitivity weighting factor matrix is ​​a 3-dimensional vector with elements of 0.85, 1.10, and 0.92, and the standardized baseline stability prediction vector is also a 3-dimensional vector with elements of 0.73, 0.64, and 0.58. The elements of the initial gain matrix generated by matching are the product of each dimension, where the first dimension is... The second dimension is The third dimension is The calculated results are 0.6205, 0.704, and 0.5336, respectively. The initial gain matrix is ​​scaled by a sensitivity amplification factor of 1.05 and a suppression factor of 0.95, resulting in coupling gain matrix elements of 0.6515, 0.6688, and 0.5070. A covariance matrix is ​​constructed from these matrices, and eigenvalue decomposition is performed. The eigenvector elements corresponding to the largest eigenvalue are 0.82, 0.45, and 0.36, which represent the dominant risk mode component. This eigenvector is used as the modal decoupling correction parameter output and applied to the subsequent nonlinear superposition of threshold curves, resulting in a significant improvement in the stability of the generated dynamic threshold core curve against high-frequency vibration disturbances.

[0178] S5.4: Based on the modal decoupling correction parameters, perform point-by-point nonlinear superposition operation on the standardized baseline stability prediction vector, introduce a hyperbolic tangent smoothing constraint mechanism to suppress threshold jumps caused by high-frequency noise, and generate a preliminary fused dynamic threshold core curve.

[0179] S5.5: Based on the statistical distribution characteristics of the preliminarily fused dynamic threshold core curve, calculate the confidence interval expansion factor, and apply the confidence interval expansion factor to the upper and lower edges of the preliminarily fused dynamic threshold core curve to generate a dynamic threshold envelope containing the upper and lower warning lines.

[0180] Based on the amplitude sequence of the initially fused dynamic threshold core curve, the amplitude set corresponding to each sampling point is read to construct a complete statistical distribution data structure. The mean and standard deviation are calculated on the statistical distribution data structure, with the mean used as the central trend position and the standard deviation as a measure of fluctuation amplitude. Based on a preset risk confidence level, the expansion factor is calculated using the normal distribution confidence interval expansion factor formula. The expansion factor is multiplied by the standard deviation to generate the interval expansion amplitude increment. The interval expansion amplitude increment is superimposed and subtracted to the mean trajectory of the initially fused dynamic threshold core curve to form an upper and lower warning line. The upper and lower warning lines are combined to form a dynamic threshold envelope data structure, and the format is standardized for subsequent boundary judgment. By superimposing the confidence interval expansion factor with the upper and lower edges of the core curve, the fusion result of the previous step is transformed into a structural stability risk assessment boundary, achieving a precise early warning threshold definition that dynamically adapts to the disturbance characteristics of the construction stage.

[0181] For example, during the rebar tying stage, the mean value of the initial fusion dynamic threshold core curve collected by on-site sensors is 0.85 mm, the standard deviation is 0.12 mm, the risk significance level α is taken as 0.05, and the corresponding standard normal distribution quantile z is 1.96. The confidence interval expansion factor is calculated according to the formula. The amplitude increment of the interval expansion is obtained. The result is 0.2352 mm. This value is superimposed on the mean trajectory of 0.85 mm to obtain the upper bound warning line. The result was 1.0852 mm; decreasing the value yielded the lower warning line. The result was 0.6148 mm. The dynamic threshold envelope was set within the range of [0.6148 mm, 1.0852 mm] during this stage. When the on-site structural state value remained within this range during multiple samplings, no warning was triggered. When the range was exceeded, it was determined to be a real stability risk, thereby significantly improving the accuracy of the warning and the environmental adaptability.

[0182] Step S6: Determine whether the measured structural state value within three consecutive sampling periods exceeds the upper or lower bound of the dynamic threshold envelope. If it does, trigger a graded early warning signal and generate a heatmap of stage matching degree and disturbance contribution. Specifically, this includes: S6.1: Perform time-series sliding window truncation processing on the measured structural state values ​​within three consecutive sampling periods to obtain a time-series sequence of measured structural state values ​​including the current time and the previous two historical times, which serves as the basic input data for subsequent threshold comparison.

[0183] Obtain the dynamic threshold envelope dataset provided in the previous sub-step and the measured structural state value of the current sampling period as the input conditions for this sub-step.

[0184] The measured structural state values ​​are arranged according to the sampling timestamp to construct the original time-series signal matrix. The matrix is ​​then truncated with a fixed step size based on the period length parameter to form three consecutive periodic data blocks covering the current time and the two previous historical times.

[0185] A sliding window function is performed on the three consecutive periodic data blocks. The window width is set to the total duration of the three periods, and the window step size is one sampling period, so as to realize the time-series slicing processing of the data blocks.

[0186] Time index normalization is applied to the slice sequence output by the sliding window function to map the time position of each period to a unified time coordinate system, ensuring accurate time alignment for subsequent threshold comparisons.

[0187] The normalized slice sequence is denoised using a Kalman filter to remove high-frequency random disturbances and retain effective signals that characterize the trend of structural state changes.

[0188] The above processing method transforms the results of the previous step into a time series of high-precision measured structural state values ​​over three consecutive sampling periods, thus enabling the construction of the basic input data for threshold envelope comparison.

[0189] For example, a high-altitude edge protection monitoring system collects structural state value sequences output by accelerometers during the concrete pouring stage, with a sampling period of 2 seconds. A window width of 6 seconds and a step size of 2 seconds are set, and the state value matrix is ​​truncated at a fixed step size, forming three periodic data blocks covering the current moment and the two previous historical moments. A sliding window calculation is applied to these data blocks, and time index normalization is performed on the data in each window to unify the time axis to the range of 0 to 6 seconds. A Kalman filter is then used, and the state transition model parameters are set as follows: The measurement noise variance is set to High-frequency noise is filtered out. The output three-period high-precision time series has a length of 3×25 data points, which serves as the input for subsequent dynamic threshold envelope out-of-bounds detection, making the calculation results of stage matching and disturbance contribution analysis more stable and accurate.

[0190] S6.2: Perform multidimensional boundary crossing detection operation based on the measured structural state value time sequence and the upper and lower bounds of the dynamic threshold envelope to generate a boundary crossing judgment flag sequence that characterizes the positional relationship of each sampling point relative to the threshold envelope.

[0191] Based on the time-series sequence of measured structural state values ​​including the current moment and the two previous historical moments, and the corresponding upper and lower bounds of the dynamic threshold envelope, a two-dimensional paired dataset with the same time index is established as the basic input for boundary crossing determination calculation. The measured value of each sampling point in the time-series sequence is differentially analyzed with the upper bound of the envelope to generate an upper bound difference sequence, which is then compared with zero to determine the over-boundary state flag. Simultaneously, the measured value of each sampling point is differentially analyzed with the lower bound of the envelope to generate a lower bound difference sequence, which is then compared with zero to determine the low-boundary state flag. A logical OR operation is used to merge the upper bound over-boundary state flag and the lower bound low-boundary state flag to form a complete set of boundary crossing state flags, covering the crossing detection requirements of both sides of the boundary. For the boundary crossing state flag set, multi-dimensional positional relationship encoding is performed, mapping each flag to the corresponding state value encoding table according to three classification labels: "above the upper bound," "below the lower bound," and "located in the interval." Using a threshold window search method, pattern recognition is performed on the flag bit sequence of continuous sampling points to generate an out-of-bounds judgment flag bit sequence that represents the positional relationship of each sampling point relative to the dynamic threshold envelope.

[0192] By using differential comparison and state coding, the results of the previous step are transformed into spatial location relationship quantification data, enabling accurate boundary judgment of measured values ​​and threshold envelopes, thereby providing high-confidence basic marker data for the logical verification of continuous boundary crossing events.

[0193] For example, in a high-altitude operation edge protection structure monitoring scenario, the sampling period is 2 seconds. The measured structural state values ​​at the current moment and the two previous moments are 0.82g, 0.79g, and 0.85g (root mean square value of vibration acceleration), respectively. The upper limits of the dynamic threshold envelope are 0.80g, 0.78g, and 0.81g, and the lower limits are 0.60g, 0.59g, and 0.61g, respectively. The difference between the current measured value of 0.82g and the upper limit of 0.80g is calculated to obtain a difference of 0.02g, which is encoded as "above the upper limit". The difference between the previous moment's 0.79g and the upper limit of 0.78g is calculated to obtain a difference of 0.01g, which is also encoded as "above the upper limit". The difference between the two previous moments' 0.85g and the upper limit of 0.81g is calculated to obtain a difference of 0.04g, which is also encoded as "above the upper limit". Similarly, using the lower bound criterion, we obtain encoding results where the values ​​at each time step are not lower than the lower bound. The upper bound out-of-bounds states at the three time steps are combined into an out-of-bounds flag sequence [1,1,1] (where 1 indicates out-of-bounds and 0 indicates not out-of-bounds). Pattern recognition is performed on this sequence, and if three consecutive sampling periods exceed the upper bound, a continuous out-of-bounds flag is output. To verify the calculation process, the absolute values ​​of the upper bound differences are aggregated to obtain the average difference. The difference is approximately 0.0233g, which is significantly higher than the preset boundary sensitivity threshold of 0.015g, proving the stability and effectiveness of the judgment result. The output boundary violation judgment flag sequence and positional relationship encoding in this scenario effectively support subsequent continuous boundary violation event screening and early warning level matching.

[0194] S6.3: Perform continuous logic AND gate verification on the out-of-bounds determination flag sequence to filter out continuous out-of-bounds event identifiers that meet the condition of breaking through the dynamic threshold envelope for three consecutive sampling periods, thereby eliminating occasional false alarms caused by instantaneous environmental disturbances.

[0195] The input conditions for performing continuous logical AND gate verification on the boundary crossing judgment flag sequence include the multi-dimensional boundary crossing detection operation results generated in step S6.2. Each detection unit contains a timestamp and a corresponding boundary crossing judgment flag. Based on these input conditions, a sliding window logical grouping operation is first performed on the flag sequence in the order of sampling periods to construct a time series group set of length three, ensuring that the time interval within the group conforms to the sensor sampling frequency calibration value. For each time series group, the continuous logical AND operation module is called to combine the three flags within the group through a bitwise AND operation. If the result is a logical 1, it indicates that there is a boundary crossing situation in all three sampling periods within the group. To ensure the rigor of the calculation, the logical AND operation result is numerically compared with a preset continuous boundary crossing judgment threshold, the threshold value being... The time-based determination is valid. For intermittent boundary violations caused by noise or data jitter, a continuous group length determination method is used based on the flag continuity verification method. This means that at least one window of continuous logic and results must satisfy the boundary violation condition to be confirmed as a continuous boundary violation event. The time series determined as continuous boundary violation events are grouped to generate continuous boundary violation event identifiers. These identifiers are bound to the corresponding sampling time range and boundary violation direction information for use by the warning level mapping module in subsequent step S6.4. Through continuous logic AND gate verification, the boundary violation determination flag sequence from the previous step is transformed into a continuous boundary violation event identifier that excludes instantaneous environmental disturbances, achieving accurate identification of continuous stability risks.

[0196] For example, in a high-altitude edge protection monitoring scenario, the sampling frequency is set to 50Hz. The structural state values ​​collected in each cycle are processed by S6.2, outputting an upper bound out-of-bounds flag sequence of [1,1,1,0,1,1,1,1,0] and a lower bound out-of-bounds flag sequence of [0,0,0,0,0,0,0,0,0]. The logic AND gate verification process performs a bitwise AND operation on the first to third cycles, resulting in... This indicates that the boundary crossing is continuous; the bitwise AND result of the 2nd to 4th period grouping is... This is not true; the bitwise AND result for the 5th to 7th period grouping is... The result is: (The result is correct); the bitwise AND operation of the 6th to 8th period groupings is: The continuous window verification requires at least one window to be valid consecutively. Therefore, the final continuous out-of-bounds event identifier covers two time ranges: periods 1-3 and periods 5-8, with the out-of-bounds direction being bounded as the upper limit for each. This processing significantly improves the identification rate of real structural risks in field verification, suppresses occasional false alarms caused by short-term surges in wind speed, and effectively improves the accuracy of early warnings.

[0197] S6.4: Based on the continuous boundary crossing event identifier and the semantic label of the current construction stage, perform a warning level mapping and matching operation to generate a graded warning signal carrying a specific risk level, so as to realize differentiated alarm output for different working conditions.

[0198] Based on the continuous boundary crossing event identifier and the semantic label of the current construction stage, the preset early warning level mapping rule library is called as the execution object. The multi-dimensional parameters of the occurrence frequency, boundary crossing range and stage disturbance source type of the continuous boundary crossing event are input into the rule matching engine to construct the mapping query conditions.

[0199] By utilizing the out-of-bounds amplitude feature parameter in the mapping query conditions, interval positioning operations are performed on the numerical classification boundaries in the rule base. By comparing the amplitude with the upper and lower limit intervals corresponding to each level boundary, a matching set of risk level candidates is identified.

[0200] By combining the frequency of continuous out-of-bounds occurrences in the mapping query conditions, a weighted scoring method is used to re-rank the candidate risk level set. The weight coefficients are dynamically adjusted according to the severity of the working conditions associated with the semantic tags of the construction stage, forming a ranked optimal risk level sequence.

[0201] Logical cross-validation is performed on the sorted risk level selection sequence to compare the consistency between the dominant disturbance source type at the current stage and the historical statistics of the candidate risk levels, and to eliminate redundant risk levels that do not conform to the actual working conditions.

[0202] The remaining risk levels are converted into graded early warning signal data packets that can be called by the early warning feedback module through an encoding mapping mechanism. The data packets contain specific risk level codes, text descriptions, and corresponding early warning output parameter configurations.

[0203] Through the above mapping matching and cross-validation processing method, the continuous out-of-bounds event identifier in the previous step is transformed into a graded early warning signal carrying a specific risk level, so as to realize differentiated alarm output for different operating conditions.

[0204] For example, during the concrete pouring stage, the peak vibration amplitude corresponding to the continuous boundary crossing event identifier is mm, frequency of occurrence The sampling period is specified by the semantic tag for the construction phase, indicating that this phase has dual disturbance attributes of high humidity operation and pump truck vibration. In the early warning level mapping rule base, the combined disturbance of high humidity operation and pump truck vibration corresponds to the risk level boundary value range of amplitude. mm and frequency ≥ Each time / cycle, the matched candidate risk level is "Level II High Risk". In the scoring method, the amplitude score is set as... Frequency score is set to According to the weighted summation formula Calculate the overall score as follows After sequential arrangement, the preferred risk level was confirmed as Level II. In cross-validation, this level showed consistency with risk statistics from similar historical periods. Therefore, the rejection rate is zero, and the final output includes the risk level code "L2", the description "Level II High Risk", and the output parameter is configured as the audible and visual alarm threshold gain multiple. The data packet. The results after execution show that the graded warning signal accurately displays the corresponding risk level and alarm method in the mobile security officer interface. In the subsequent generation of the S6.5 stage matching degree and disturbance contribution heatmap, the risk level information and the quantification results of the deviation degree remain consistent, thus significantly improving the targeting and operability of the alarm.

[0205] S6.5: Using the graded early warning signal to trigger the stage matching degree calculation module and the disturbance contribution analysis module, the deviation of the current measured structural state value from the benchmark stability decay trajectory template and residual compensation coefficient is quantitatively decomposed to generate a heat map that intuitively displays the stage matching degree and disturbance contribution of the risk sources.

[0206] Based on the graded early warning signal as the triggering condition, the stage matching degree calculation module is invoked to obtain the correlation strength between the semantic label of the current construction stage and the benchmark stability decay trajectory template.

[0207] The difference between the measured structural state value and the predicted reference value of the baseline stability decay trajectory template at the same time stamp is calculated to obtain the structural response deviation vector. The sum of squares of the deviation vector is then performed to quantify the overall deviation.

[0208] A normalization process is used to map the magnitude domain of the deviation vector to a dimensionless interval, calculate the stage matching coefficient, and output it to the visualization engine.

[0209] The disturbance contribution analysis module is invoked to decompose the residual compensation coefficients hierarchically according to the disturbance source classification labels. The cumulative summation and proportion calculation are performed on the contribution value of each disturbance source to generate a disturbance contribution rate vector.

[0210] Based on the preset color mapping rules, the stage matching degree coefficient is used as the vertical index of the heat map, and the disturbance contribution rate vector is used as the horizontal index of the heat map. A two-dimensional heat map matrix is ​​generated by matrix filling and weighted mapping to form a composite heat map of stage matching degree and disturbance contribution rate.

[0211] Through the above chain processing method, the graded early warning signals from the previous step are transformed into visual analysis data that can intuitively reflect the composition of risk sources, so as to achieve the expected technical effect of safety officers quickly judging the actual structural risks and inherent disturbances of the stage.

[0212] For example, during the concrete pouring stage of a high-rise building, the upper limit of the dynamic threshold envelope is 3.5mm displacement, and the lower limit is -3.2mm displacement. The measured structural state values ​​over three consecutive sampling periods are 4.2mm, 4.0mm, and 3.8mm, respectively, triggering a level-three warning signal. The stage matching degree calculation module calls the template prediction value of 3.6mm, with a difference vector of [0.6, 0.4, 0.2]mm. After performing square sum normalization, an amplitude of 0.072 is obtained. This is inserted into the Pearson correlation coefficient formula, where n=3, x is the measured value sequence, and y is the template value sequence. The calculated stage matching degree coefficient is 0.94, mapped to a dark red color code. The disturbance contribution analysis module extracts the residual compensation coefficients according to the disturbance source classification: vibration source contribution is 0.015, temperature and humidity contribution is 0.008, and load change contribution is 0.005. The cumulative contribution rate vector is [0.015, 0.008, 0.005], and the proportion vector is [0.50, 0.27, 0.23]. In the heat map matrix, the vertical axis takes a stage matching degree of 0.94, corresponding to a dark red color, and the horizontal axis takes the contribution proportion, mapping different color intensities to different disturbance sources. The output synthetic heat map intuitively shows that vibration source is the main risk factor, followed by temperature and humidity, and load change is the weakest, enabling safety officers to quickly identify the causes of risky structures and adjust on-site protection strategies on mobile devices.

[0213] Step S7: Receive the on-site handling records and threshold execution effect feedback data after the early warning event closure, and inject the data including stage labels, false alarm / missed alarm identifiers, and response delay duration into the stage perturbation knowledge graph. Specifically, this includes: S7.1: Obtain on-site handling records and system log data after the closed loop of the early warning event, and use natural language processing technology to extract entities and perform semantic mapping on the unstructured text to generate a structured feedback dataset containing semantic labels of standard construction stages, binary false alarm and missed alarm identifiers, and quantitative response delay duration values.

[0214] S7.2: Based on the standard construction stage semantic labels in the structured feedback dataset, perform node localization operations in the stage disturbance knowledge graph, retrieve and extract existing disturbance source type nodes directly associated with the label and their corresponding physical constraint attribute edges, and form a subgraph context to be updated.

[0215] S7.3: Using the binary false positive and false negative identifier in the context of the subgraph to be updated as a confidence correction factor, and combining it with the quantized response delay duration to construct a dynamic penalty function, the initial association weights of each disturbance source type node in the context of the subgraph to be updated are nonlinearly recalculated to generate intermediate state graph data containing weight correction values.

[0216] A confidence correction mechanism is constructed based on the binary false positive / false negative identifiers and quantized response delay duration values ​​in the context of the subgraph to be updated. The input objects are the perturbation source type nodes and their initial association weights located in the previous sub-step. The binary false positive / false negative identifiers are mapped to different confidence decay coefficients according to the two categories of false positives and false negatives. False positives correspond to reducing association confidence to reduce the impact of interfering features, while false negatives correspond to increasing association confidence and compensating for uncaptured risk features. The quantized response delay duration is normalized and mapped to a delay factor value within a standard range. The delay factor is used to characterize the influence of the warning lag on the model association. A dynamic penalty function is established between the decay coefficients and the delay factor, and a nonlinear combination method is used to achieve the coupling adjustment of the two. A nonlinear recalculation operation is performed on the output of the above dynamic penalty function, and the adjusted weights are scaled to avoid the impact of instantaneous jumps on the stability of the graph structure, and intermediate graph data containing weight correction values ​​are generated. By using a dynamic penalty function and a smooth adjustment process, the localization results from the previous step are transformed into weight correction data that can be used for subsequent incremental updates of the map topology, thus achieving the expected technical effect of refined optimization of association strength.

[0217] For example, in a case of monitoring edge protection during high-altitude operations, the subgraph context to be updated contains 5 nodes of disturbance source types, with initial association weights of 0.65, 0.78, 0.55, 0.62, and 0.81, respectively; among them, 2 nodes correspond to false alarms in the field, with the false alarm code value set to 1, and the mapping attenuation coefficient... The value is set to 0.08. The other three nodes correspond to missed detections, with the missed detection code set to 0. The mapping attenuation coefficient is negative -0.05 for weighted compensation. The quantization response delay times are 4.2s, 3.8s, 5.1s, 4.7s, and 3.5s, respectively. After normalization to the interval [0,1], the corresponding delay factor β values ​​are 0.042, 0.038, 0.051, 0.047, and 0.035. Substituting the above parameters into the dynamic penalty function: for example, the first node calculates the adjusted weight as... The output adjustment value is approximately 0.598. After performing the same calculation on all nodes, the corrected weights are concentrated in the range of [0.59, 0.85]. After smoothing, the impact of sudden increases or decreases on the system's topological stability is avoided. This result is used as intermediate graph data for subsequent incremental updates, effectively improving the accuracy of knowledge graph association weights and the ability to adapt to dynamic disturbances at similar stages.

[0218] S7.4: Based on the weight correction values ​​in the intermediate state graph data, perform an incremental update operation on the graph topology structure, embed historical cases containing standard construction stage semantic tags, binary false alarm and false alarm identifiers, and quantified response delay duration values ​​as new evidence nodes into the stage perturbation knowledge graph, and generate a reverse injection enhanced knowledge graph with temporal evolution characteristics.

[0219] Based on the weight correction values ​​in the intermediate state graph data, incremental update operations are performed on the existing topology of the stage-perturbation knowledge graph to ensure the correct embedding of new evidence nodes and structural coherence. The standard construction stage semantic labels in the structured feedback dataset are used as the primary index, establishing unique binding relationships in the graph node list, and setting the insertion order of the label node to construct a temporal evolution chain. The binary false positive / false negative identifiers are mapped to the direction of associated weight adjustment based on their positive or negative values. The weight update interface is called to multiply this direction with the intermediate state weight correction value, forming the final update coefficient array of node edge weights. The quantized response delay duration is converted into a time weight decay factor, which is multiplied and applied to the edge weights corresponding to the update coefficient array to form a set of updated edges with temporal decay attributes. Historical case nodes containing standard construction stage semantic labels, false positive / false negative identifiers, and response delay duration values ​​are constructed, and a timestamp field is added to the node attribute table to identify the absolute time of the case occurrence, ensuring its retrieval on the graph timeline. Based on the physical constraint attributes between stage nodes and disturbance source type nodes, historical case nodes are embedded into the existing topology by creating new edges. The metadata field of each edge records the corrected weight value and time weight decay factor derived from the updated edge set, thus dynamically representing the association strength between historical cases and existing nodes. This process transforms the result of the previous step into a reverse-injection enhanced knowledge graph with temporal evolution characteristics, enabling continuous adaptive updates of the knowledge graph to the evolutionary patterns of disturbances during the construction phase.

[0220] For example, in the monitoring of high-altitude operations during the rebar tying stage, the weight correction value in the intermediate state map data is set to 0.15, the standard construction stage semantic label is "rebar tying," the binary false alarm / false negative identifier is 1 (indicating a false alarm), and the quantized response delay is 12.5 seconds. The false alarm identifier is mapped to the weight reduction direction (negative adjustment), and multiplied by the weight correction value of 0.15 to obtain the update coefficient. Convert the delay duration into a time-weighted decay factor. The attenuation coefficient is obtained. The final edge weight correction value is obtained by multiplying the decay coefficient by the update coefficient. This information is recorded in the metadata of the newly created edge between the case node and the disturbance source type node. The historical case node also embeds the timestamp "2024-03-15 10:32:21" and establishes a unique binding relationship with the stage node "reinforcement binding". In subsequent graph retrieval, this back-injected enhanced knowledge graph can significantly improve the accuracy of identifying long-term vibration disturbance patterns in the reinforcement binding stage and reduce false alarms caused by fixed thresholds.

[0221] S7.5: Perform consistency verification and redundancy pruning on the back-injected enhanced knowledge graph, and output the final updated stage disturbance knowledge graph instance to ensure that the instance can accurately reflect the latest statistical laws of disturbance source type and physical constraint attributes under the current construction stage, so as to be used for subsequent baseline stability decay trajectory template fitting accuracy optimization.

[0222] Step S8: Based on the injected feedback data, the association weights in the stage perturbation knowledge graph are iteratively updated to optimize the fitting accuracy of the baseline stability decay trajectory template under subsequent similar construction stages, thereby completing the adaptive adjustment closed loop. Specifically, this includes: S8.1: Obtain a feedback data packet containing stage labels, false alarm / missed alarm identifiers, and response delay duration. Use natural language processing technology to perform semantic parsing on the on-site handling record text to extract structured handling action feature vectors. Then, perform spatiotemporal alignment and fusion of the structured handling action feature vectors with the feedback data packet to generate a composite feedback sample set with weight correction labels.

[0223] S8.2: Based on the composite feedback sample set with weight correction labels, the gradient deviation value of the associated weights of each node in the current stage of the perturbation knowledge graph is calculated using the error backpropagation algorithm. The gradient deviation value is then mapped to the dynamic adjustment coefficient of the graph edge weights to generate the associated weight increment matrix to be updated.

[0224] S8.3: The original association weights in the stage perturbation knowledge graph are updated using the incremental association weight matrix to be updated. This process is nonlinearly weighted and updated to eliminate the distortion of the graph topology caused by the identification bias of environmental perturbation, thereby generating an optimized stage perturbation knowledge graph with self-evolution capabilities.

[0225] S8.4: Based on the optimized stage perturbation knowledge graph with self-evolution capability, the time series curve clusters in the historical building engineering monitoring database are re-retrieved and matched, and the time series curve clusters are re-ranked by similarity using the weighted dynamic time warping method to select a set of preferred time series curves with high confidence.

[0226] S8.5: Perform normalization aggregation operation on the set of preferred time-series curves with high confidence, reconstruct the benchmark stability decay trajectory template that characterizes the structural response law of the construction stage, and replace the original template with the reconstructed benchmark stability decay trajectory template to improve the fitting accuracy of the subsequent dynamic threshold envelope.

[0227] The intelligent monitoring method for the stability of edge protection in high-altitude operations of building engineering provided in this application can be applied to, for example... Figure 4 In the application environment shown, terminal 102 communicates with server 104 via a network. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices, and server 104 can be a standalone server or a server cluster consisting of multiple servers.

[0228] In one embodiment, an intelligent monitoring device for the stability of edge protection during high-altitude operations in construction engineering is provided, comprising: a semantic tag generation module, a feature extraction module, a template construction module, a compensation calculation module, a threshold generation module, an early warning module, a feedback module, and an update module, wherein: Semantic tag generation module: acquires the building information model progress data stream and on-site positioning beacon data, and generates construction stage semantic tags carrying spatiotemporal attributes based on time window alignment and spatial coordinate matching; Feature extraction module: Based on the semantic tags of the construction stage, retrieve the knowledge graph of stage disturbances, extract the dominant disturbance source type and physical constraint attributes under the current working condition to generate stage disturbance feature vectors; Template construction module: Utilizes stage disturbance feature vector retrieval and matching to retrieve parameter time series curves of similar projects with similar current working conditions from the historical building engineering monitoring database, and constructs a benchmark stability decay trajectory template that characterizes the structural response law of the construction stage after normalization; The calculation and compensation module acquires real-time sensor data streams and calculates vibration energy entropy, displacement trend slope, and temperature and humidity coupling offset, which are then used as short-term disturbance input conditions to input the long short-term memory network to output residual compensation coefficients. Threshold generation module: Performs nonlinear fusion calculations on the predicted values ​​of the baseline stability decay trajectory template and the residual compensation coefficients to generate a dynamic threshold envelope for the current construction stage; Early warning module: Determines whether the measured structural state value within three consecutive sampling periods exceeds the upper or lower bound of the dynamic threshold envelope. If it does, it triggers a graded early warning signal and generates a heat map of stage matching degree and disturbance contribution. Feedback module: Receives on-site handling records and threshold execution effect feedback data after the early warning event is closed, and injects data including stage labels, false alarm and missed alarm indicators and response delay duration into the stage perturbation knowledge graph; Update module: Based on the injected feedback data, iteratively update the association weights in the stage disturbance knowledge graph, optimize the fitting accuracy of the baseline stability decay trajectory template under subsequent similar construction stages, and complete the adaptive adjustment closed loop.

[0229] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an intelligent monitoring method for the stability of edge protection in high-altitude construction projects.

[0230] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: S1: Obtain the building information model progress data stream and on-site location beacon data, and generate construction stage semantic tags carrying spatiotemporal attributes based on time window alignment and spatial coordinate matching.

[0231] S2: Based on the semantic tags of the construction stage, retrieve the knowledge graph of stage disturbances, extract the dominant disturbance source type and physical constraint attributes under the current working condition to generate stage disturbance feature vectors.

[0232] S3: Utilize the stage disturbance feature vector retrieval and matching to retrieve the parameter time series curves of similar projects with the current working conditions from the historical building engineering monitoring database. After normalization, construct a benchmark stability decay trajectory template that characterizes the structural response law of the construction stage.

[0233] S4: Acquire real-time sensor data streams and calculate vibration energy entropy, displacement trend slope, and temperature and humidity coupling offset, and input these as short-term disturbance input conditions into the long short-term memory network to output residual compensation coefficients.

[0234] S5: Perform nonlinear fusion calculations on the predicted values ​​of the baseline stability decay trajectory template and the residual compensation coefficients to generate a dynamic threshold envelope for the current construction stage.

[0235] S6: Determine whether the measured structural state value within three consecutive sampling periods exceeds the upper or lower bound of the dynamic threshold envelope. If it does, trigger a graded early warning signal and generate a heat map of stage matching degree and disturbance contribution.

[0236] S7: Receives on-site handling records and threshold execution effect feedback data after the early warning event is closed, and injects data including stage labels, false alarm and missed alarm indicators and response delay duration into the stage disturbance knowledge graph.

[0237] S8: Based on the injected feedback data, the association weights in the stage disturbance knowledge graph are iteratively updated to optimize the fitting accuracy of the baseline stability decay trajectory template under the same construction stage in order to complete the adaptive adjustment closed loop.

[0238] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0239] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0240] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0241] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent monitoring of the stability of edge protection during high-altitude operations in construction engineering, specifically comprising: S1: Obtain the building information model progress data stream and on-site positioning beacon data, generate construction stage semantic tags with spatiotemporal attributes based on time window alignment and spatial coordinate matching, and construct a stage disturbance knowledge graph containing work condition root nodes, disturbance source type nodes and physical constraint attribute nodes. S2: Based on the semantic tags of the construction stage, retrieve the knowledge graph of stage disturbances, extract the dominant disturbance source type and physical constraint attributes under the current working condition to generate stage disturbance feature vectors; S3: Use the stage disturbance feature vector to retrieve the parameter time series curves of similar projects with similar current working conditions from the historical building engineering monitoring database, and construct a benchmark stability decay trajectory template that characterizes the structural response law of the construction stage after normalization. S4: Acquire real-time sensor data streams and calculate vibration energy entropy, displacement trend slope and temperature and humidity coupling offset, and input them as short-term disturbance input conditions into the long short-term memory network to output residual compensation coefficients. S5: Perform nonlinear fusion calculations on the predicted values ​​of the baseline stability decay trajectory template and the residual compensation coefficients to generate a dynamic threshold envelope for the current construction stage. S6: Determine whether the measured structural state value within three consecutive sampling periods exceeds the upper or lower bound of the dynamic threshold envelope. If it does, trigger a graded early warning signal and generate a heat map of stage matching degree and disturbance contribution.

2. The intelligent monitoring method for the stability of edge protection in high-altitude operations of building engineering according to claim 1, characterized in that, Raw monitoring data from on-site monitoring systems of historical engineering projects are collected, and after data cleaning, amplitude normalization, and time axis alignment, a historical building engineering monitoring database containing time-series curves and their associated engineering topology parameters and physical constraint attributes is constructed.

3. The intelligent monitoring method for the stability of edge protection in high-altitude operations of building engineering according to claim 1, characterized in that, Following S6, the following is also included: S7: Receive the on-site handling records and threshold execution effect feedback data after the early warning event is closed, and inject the data including stage labels, false alarm and missed alarm indicators and response delay duration into the stage perturbation knowledge graph. S8: Based on the injected feedback data, the association weights in the stage disturbance knowledge graph are iteratively updated to optimize the fitting accuracy of the baseline stability decay trajectory template under the same construction stage in order to complete the adaptive adjustment closed loop.

4. The intelligent monitoring method for the stability of edge protection in high-altitude operations of building engineering according to claim 1, characterized in that, Step S3 specifically includes: Based on the historical building engineering monitoring database and stage disturbance feature vectors, the Euclidean distance score between each historical time series curve and the current working condition features is calculated, and a cluster of high-confidence candidate time series curves with similar topological structure and consistent physical properties is selected. For each original time series curve in the high-confidence candidate time series curve cluster, time axis resampling is performed to eliminate dimensional differences and generate a standardized historical response sequence set; Dynamic time warping and alignment are performed based on a standardized historical response sequence set to identify the characteristic peak moments of key inflection points representing structural stability in each sequence, correct time offset errors, and generate a family of phase-aligned time series curves. Perform point-by-point statistical aggregation on the time series curve family, calculate the arithmetic mean of the amplitudes of all curves at each time slice point as the central trend term, and calculate the standard deviation as the discrete fluctuation term to generate a preliminary statistical template containing the mean trajectory line and the variance envelope. Local high-frequency jitter is removed from the mean trajectory line, and the variance envelope is expanded and corrected by combining the confidence coefficient, and a baseline stability decay trajectory template is output.

5. The intelligent monitoring method for the stability of edge protection in high-altitude operations of building engineering according to claim 1, characterized in that, Step S4 specifically includes: The original time-series signals of the accelerometer and displacement sensor are acquired. Segmented acceleration data blocks are generated based on the original time-series signals of the accelerometer, and the frequency domain power spectral density distribution is extracted. Then, disorder quantification is performed to obtain the vibration energy entropy value. The slope of the displacement trend is generated based on the original time-series signal from the displacement sensor. Ambient temperature and relative humidity data are acquired and normalized to generate standardized temperature and humidity variables. The temperature and humidity coupling strength is quantified based on the standardized temperature and humidity variables. Then, the temperature and humidity coupling offset is obtained by weighted fusion of the temperature and humidity coupling strength in combination with the corrosion acceleration factor mapping table. The vibration energy entropy, displacement trend slope and temperature and humidity coupling offset are aligned and stitched together according to the timestamp to construct a multidimensional short-term disturbance feature vector, and tensor reshaping is performed on it to generate three-dimensional tensor data. Three-dimensional tensor data is input into a long short-term memory network model to generate cell state update quantities. The hidden layer state features are extracted, and linear regression prediction is performed on the hidden layer state features through a fully connected layer to output scalar form residual compensation coefficients.

6. The intelligent monitoring method for the stability of edge protection in high-altitude operations of building engineering according to claim 5, characterized in that, The method of generating the displacement trend slope based on the original time-series signal of the displacement sensor specifically involves acquiring the original time-series signal of the displacement sensor within the same time window, performing a first-order difference operation on the original time-series signal of the displacement sensor to generate a displacement change rate sequence, and performing least squares linear fitting processing on the displacement change rate sequence to extract the linear regression slope parameter, thereby determining the linear regression slope parameter as the displacement trend slope characterizing the structural deformation trend.

7. The intelligent monitoring method for the stability of edge protection in high-altitude operations of building engineering according to claim 1, characterized in that, Step S5 specifically includes: Based on the output node of the baseline stability decay trajectory template at the current moment, the baseline stability prediction amplitude representing the natural evolution trend of the structure is extracted, and its dimensions are normalized to generate a standardized baseline stability prediction vector. By utilizing the residual compensation coefficients output by the Long Short-Term Memory Network and combining them with a family of nonlinear activation functions, dynamic weight mapping is performed on the residual compensation coefficients to generate a perturbation sensitivity weighting factor. Based on the corresponding dimensional relationship between the perturbation sensitivity weighting factor and the baseline stability prediction vector, a coupling gain matrix is ​​constructed, and eigenvalue decomposition is performed on it to generate modal decoupling correction parameters. The baseline stability prediction vector is nonlinearly superimposed point by point based on the modal decoupling correction parameters. A hyperbolic tangent smoothing constraint mechanism is introduced to suppress threshold jumps caused by high-frequency noise, and a preliminary fusion dynamic threshold core curve is generated. Based on the statistical distribution characteristics of the dynamic threshold core curve, the confidence interval expansion factor is calculated and applied to the upper and lower edges of the initially fused dynamic threshold core curve to generate a dynamic threshold envelope containing the upper and lower warning lines.

8. An intelligent monitoring device for the stability of edge protection during high-altitude operations in construction engineering, characterized in that, The device includes: Semantic tag generation module: acquires the building information model progress data stream and on-site location beacon data, and generates semantic tags for the construction phase; Feature extraction module: Retrieves a knowledge graph of stage perturbations based on semantic tags of construction stages, and generates stage perturbation feature vectors; Template construction module: Use stage perturbation feature vectors to retrieve time series curves from the historical building engineering monitoring database to construct a baseline stability decay trajectory template; The calculation and compensation module acquires real-time sensor data streams and calculates vibration energy entropy, displacement trend slope, and temperature and humidity coupling offset, which are then input into a long short-term memory network to output residual compensation coefficients. Threshold generation module: Generates a dynamic threshold envelope based on the baseline stability decay trajectory template and residual compensation coefficient; Early warning module: If the measured structural state value exceeds the upper or lower bound of the dynamic threshold envelope within three consecutive sampling periods, a graded early warning signal is triggered and a heat map of stage matching degree and disturbance contribution is generated. Feedback module: Receives on-site handling records and threshold execution effect feedback data, and injects the data into the stage perturbation knowledge graph; Update module: Iteratively updates the knowledge graph of stage perturbations based on feedback data.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that... When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1-7.

Images