Stress-sensing based disc-buckle type scaffold safety warning system and method

Abstract

This invention relates to a stress-sensing-based safety early warning system and method for disc-lock scaffolding, belonging to the field of construction safety monitoring technology. The system includes a construction stage sensing module to acquire information about the current construction stage, an early warning effectiveness evaluation module to generate effectiveness evaluation data based on stress sensor data, an adaptive optimization decision-making module to output an early warning threshold adjustment strategy through a reinforcement learning model, a dynamic early warning threshold execution module to update the threshold in real time, and an early warning alarm module to compare real-time data with the updated threshold to trigger an early warning signal. This invention enables dynamic optimization of the early warning threshold based on the construction stage and real-time data, avoiding false alarms or missed alarms caused by fixed thresholds.

Description

Technical Field

[0001] This invention relates to the field of construction safety monitoring technology, specifically to a safety early warning system and method for disc-lock scaffolding based on stress sensing. Background Technology

[0002] Construction safety monitoring technology is a crucial link in ensuring the structural stability of large-scale engineering projects. In the construction of high-rise buildings and complex structures, the safety of modular scaffolding as a temporary support system directly affects the overall project progress and personnel safety. Current technologies for modular scaffolding safety early warning systems generally employ a fixed stress threshold mechanism. This involves real-time data collection via stress sensors deployed at key nodes, comparing the collected values ​​with preset static thresholds, and triggering an alarm signal when the threshold is exceeded. This system typically combines a construction schedule with manually set threshold parameters, relies on historical experience data to determine the alarm critical point, and uses a central control unit to aggregate sensor information for centralized monitoring.

[0003] However, the fixed threshold warning method is difficult to accurately match the actual load-bearing state of the structure during the dynamic changes in the construction phase. Summary of the Invention

[0004] This invention provides a safety early warning system and method for disc-lock scaffolding based on stress sensing, which can solve the problem of false alarms or missed alarms caused by the inability of traditional early warning systems to adapt to the dynamic changes in load of disc-lock scaffolding at different construction stages due to fixed stress thresholds.

[0005] This invention provides a stress-sensing-based disc-lock scaffolding safety early warning system, comprising: The construction phase perception module is used to acquire and determine the current construction phase information of the disc-lock scaffolding; The early warning effectiveness evaluation module is communicatively connected to stress sensors deployed at key nodes of the disc-lock scaffolding. It is used to generate effectiveness evaluation data for the current early warning threshold based on historical and real-time stress data and early warning history data collected by the stress sensors. The adaptive optimization decision module has a built-in reinforcement learning model. The reinforcement learning model takes the current construction stage information and the performance evaluation data as state inputs, and after calculation by the decision network, outputs the early warning threshold adjustment strategy for different monitoring areas. The early warning threshold dynamic execution module is connected to the adaptive optimization decision module and the early warning alarm module respectively, and is used to adjust the strategy according to the early warning threshold and update the corresponding early warning threshold in the early warning alarm module in real time. The early warning alarm module is used to compare the real-time stress data collected by the stress sensor with the updated early warning threshold. If the threshold is exceeded, a safety early warning signal is triggered.

[0006] Preferably, the construction phase sensing module further includes: The visual progress recognition unit is used to collect image data through cameras deployed at the construction site and to identify the appearance, movement and distribution of key construction components based on computer vision algorithms. The digital twin automatic update unit is connected to the visual progress recognition unit and is used to automatically drive the component status and progress plan labels in the BIM digital twin model to be updated synchronously according to the recognition results. The current construction phase information is automatically determined by the BIM digital twin model based on the updated progress status.

[0007] Preferably, the early warning effectiveness evaluation module further integrates: The causal inference analysis unit is used to construct a causal graph model based on historical stress data and alarm records, and to quantitatively analyze the strength of the causal relationship between specific construction activities, environmental changes and abnormal stress fluctuations. The Bayesian dynamic evaluation unit receives the output of the causal inference analysis unit and constructs the performance evaluation data into a probability distribution that changes over time. The prior probability is set based on the historical false alarm / false alarm rate, and the posterior probability is dynamically updated based on the real-time alarm results and on-site verification feedback.

[0008] Preferably, the reinforcement learning model in the adaptive optimization decision module is a multi-agent reinforcement learning architecture, wherein: The first intelligent agent makes decisions with the goal of maximizing the overall structural safety factor. The second intelligent agent makes decisions with the goal of minimizing interference with construction efficiency. Two intelligent agents negotiate strategies and conduct collaborative training through a cooperative game framework in game theory. The early warning threshold adjustment strategy is the Pareto optimal solution negotiated by both parties.

[0009] Preferably, the early warning threshold dynamic execution module further includes: The federated learning aggregation unit is used to periodically receive locally optimized threshold adjustment model parameters from cloud edge gateways of multiple similar projects while ensuring the data privacy of each construction site, and to perform secure aggregation to generate a globally optimized model. The dynamic threshold simulation and deduction unit, based on the aggregated model and the current digital twin of the disc-lock scaffolding, performs short-term stress response simulation on the threshold to be adjusted, predicts the alarm modes that may be triggered after the adjustment, and if the simulation shows that it may cause an alarm storm or cover up the real risk, it triggers the adjustment strategy rollback and manual review request.

[0010] Preferably, the early warning alarm module further includes: The multimodal fusion alarm unit receives simultaneously a data stream from a stress sensor, a vibration / tilt data stream from an environmental sensor, and a real-time image stream from the visual progress recognition unit described in claim 2 at its input terminal. An attention-based alarm decision engine is used to perform weighted fusion analysis on the above multimodal data streams. The highest level of deterministic alarm is triggered only when evidence from multiple modalities points to the same high-risk event. If only a single modality is abnormal, a lower level of suggestive check alarm is triggered.

[0011] Preferably, the multi-agent reinforcement learning architecture also integrates: The swarm intelligence optimizer uses the ant colony algorithm to simulate multiple virtual agents exploring in parallel in the policy space. Each agent uses the initial negotiation results of the cooperative game framework as the initial pheromone distribution to perform a local fine-grained search. The dynamic environment adaptability assessment unit constructs a dynamic environment complexity index based on real-time meteorological data, equipment operating status, and worker behavior pattern data. When the index exceeds a threshold, it automatically increases the exploration weight of the swarm intelligent optimizer to discover emergency strategies to adapt to sudden complex situations.

[0012] Preferably, the cooperative game framework further includes: The social sentiment computing module uses sentiment computing algorithms to analyze the decision preference patterns of construction site managers in the trade-off between safety and efficiency, based on their historical decision-making data. The strategy explanation and visualization unit transforms the decision-making logic, trade-offs, and rationality of the final strategy of each agent during the negotiation process into an understandable security briefing using natural language generation technology, and then visualizes it using augmented reality devices to help managers understand and trust the system's decision-making.

[0013] Preferably, the federated learning aggregation unit further includes: The anomaly model detection and defense module uses adversarial machine learning technology to perform feature analysis on the local models submitted by each construction site before parameter aggregation, and to detect and isolate any malicious or low-quality model updates. The adaptive weighted aggregation optimizer dynamically adjusts the weight of each construction site in the federated aggregation based on its data quality score, model update stability index, and historical contribution. The data quality score is calculated by combining the output of the anomaly model detection and defense module, the construction site supervision score, and the historical early warning accuracy.

[0014] The present invention also provides a safety early warning method for disc-lock scaffolding based on stress sensing, and uses the aforementioned safety early warning system for disc-lock scaffolding based on stress sensing to realize safety early warning for disc-lock scaffolding.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention obtains information about the current construction stage through a construction stage perception module, enabling the system to identify different working conditions such as the concrete pouring period or the formwork removal and cleaning period, thereby providing key contextual basis for threshold adjustment. This invention generates dynamic performance evaluation data based on the comprehensive analysis of historical and real-time stress data and early warning history records by the early warning performance evaluation module, enabling the system to quantify the applicability of the current threshold. This invention utilizes the reinforcement learning model built into the adaptive optimization decision module to perform decision network calculations using construction phase information and performance evaluation data as state inputs, and outputs early warning threshold adjustment strategies for different monitoring areas, enabling the thresholds to be autonomously optimized as the construction process progresses. This invention updates the threshold parameters in the early warning alarm module in real time through the early warning threshold dynamic execution module, ensuring that the threshold adjustment strategy takes effect immediately; This invention utilizes an early warning alarm module to compare real-time stress data with an updated threshold to trigger an early warning signal, thereby achieving a precise response to the structural safety status. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a flowchart of the construction phase sensing module in this invention; Figure 3 This is a flowchart illustrating the further integration of the early warning effectiveness evaluation module in this invention. Detailed Implementation

[0017] refer to Figures 1 to 3 In the safety management of modular scaffolding construction, a common problem is that early warning systems rely on fixed stress thresholds. Because the types, distribution characteristics, and structural stiffness responses of loads on modular scaffolding vary at different construction stages, using a uniform alarm threshold can easily lead to frequent false alarms during peak concrete pouring periods, or to missed alarms due to insufficient sensitivity during the unloaded formwork removal phase. While existing technologies have deployed stress sensors and possess basic alarm functions, their early warning logic is not dynamically coupled with the construction progress and lacks a quantitative feedback mechanism for historical early warning effects, making it difficult to support continuous optimization of threshold parameters as working conditions evolve.

[0018] To address the aforementioned problems, this invention proposes a stress-sensing-based safety early warning system and method for disc-lock scaffolding. The system of this invention includes: The construction phase perception module is used to acquire and determine the current construction phase information of the disc-lock scaffolding; The early warning effectiveness assessment module communicates with stress sensors deployed at key nodes of the disc-lock scaffolding. It is used to generate effectiveness assessment data for the current early warning threshold based on historical and real-time stress data and early warning history data collected by the stress sensors. The adaptive optimization decision module has a built-in reinforcement learning model. The reinforcement learning model takes the current construction stage information and performance evaluation data as state input, and after calculation by the decision network, it outputs the early warning threshold adjustment strategy for different monitoring areas. The early warning threshold dynamic execution module is connected to the adaptive optimization decision module and the early warning alarm module respectively. It is used to adjust the strategy according to the early warning threshold and update the corresponding early warning threshold in the early warning alarm module in real time. The early warning and alarm module is used to compare the real-time stress data collected by the stress sensor with the updated early warning threshold. If the threshold is exceeded, a safety early warning signal is triggered.

[0019] The construction phase perception module identifies the current construction progress through multi-source information fusion. Its inputs include the BIM schedule timeline, on-site image recognition results, structured fields of the construction log, or equipment operation status signals. This construction phase perception module is a software module deployed in an edge computing gateway or cloud server. Its output is a structured phase identifier, such as the third layer of concrete pouring, area A of the masonry work surface, or the formwork removal preparation stage. This structured phase identifier can be set according to the actual situation. For example, it can be a discrete phase code or a continuous progress percentage. This embodiment of the invention does not impose any special limitations on this.

[0020] The early warning effectiveness evaluation module quantifies and analyzes the actual performance of the current early warning threshold. Its inputs include time-series stress data streams collected by stress sensors, records of early warning events triggered within the corresponding time period, and labels of on-site manual verification results. The early warning effectiveness evaluation module generates effectiveness evaluation data based on statistical deviation analysis, alarm hit rate sliding window calculation, or false alarm / missed alarm weighted index modeling. For example, it can be expressed as a confidence score between 0 and 1, normal distribution parameters, or Bayesian posterior probability density function. The early warning effectiveness evaluation data can be set according to the actual situation. For example, it can be a scalar value, a vector, or a probability distribution. This embodiment of the invention does not impose any special limitations on this.

[0021] The adaptive optimization decision module embeds a reinforcement learning model and executes threshold policy generation. Its built-in reinforcement learning model is a deep Q-network (DQN), proximal policy optimization (PPO), or an Actor-Critic architecture. The state space of the reinforcement learning model contains a joint representation of the current construction stage information and performance evaluation data. The action space is defined as the threshold offset applied to each monitoring area of ​​the disc-lock scaffold. The reward function comprehensively considers the penalty for false alarms, the penalty for missed alarms, and the positive incentive for the pass rate of expert manual review. The training data of the reinforcement learning model comes from the historical project multi-stage stress-alarm-review closed-loop dataset. Its model parameters can be set according to the actual situation. For example, they can be pre-trained weights or online fine-tuned parameters. This embodiment of the invention does not impose any special limitations on this.

[0022] The dynamic execution module for early warning thresholds enables real-time distribution and application of threshold parameters. Its input is the early warning threshold adjustment strategy, and its output is the updated threshold value mapped to each sensor channel or area group. This module supports millisecond-level threshold refresh and can be completed using a publish-subscribe message mechanism or direct register writing. Its execution process is configured with a rollback mechanism, which automatically restores to the previous stable version when a sudden policy adjustment triggers continuous abnormal alarms. The communication protocol of this dynamic execution module can be set according to actual conditions, such as MQTT, ModbusTCP, or a private binary protocol. This embodiment of the invention does not impose any special limitations on this.

[0023] The early warning alarm module performs final criterion comparison and signal triggering. Its inputs include real-time sampled values ​​from the stress sensor, updated regionalized early warning thresholds, and alarm suppression time window configurations. The early warning alarm module performs point-by-point comparison operations. When the stress value at any monitoring point exceeds the corresponding threshold and the duration is greater than the set hysteresis time, a graded alarm signal is triggered. The alarm signal can be a digital IO level, a CAN bus message, or an HTTP callback request. Its output interface can be set according to actual conditions. For example, it can be a sound and light alarm drive signal, a mobile terminal push message, or a BIM platform alarm layer mark. This embodiment of the invention does not impose any special limitations on this.

[0024] The core innovation of this invention lies in constructing a five-order coupled adaptive early warning control link: construction phase—performance feedback—reinforcement decision-making—dynamic execution—closed-loop alarm. This upgrades the traditional open-loop static threshold system into a closed-loop intelligent agent system with state perception, effect evaluation, autonomous decision-making, and execution feedback capabilities. Construction phase information serves as a priori working condition guidance signal, and performance evaluation data serves as a posterior effect calibration signal. Together, they constitute the state input for reinforcement learning, enabling the model to autonomously converge to a threshold strategy distribution that balances safety and robustness under different load evolution paths.

[0025] The working process and principle of this invention are as follows: After the system starts, the construction stage perception module continuously identifies the current construction stage and synchronizes it to the early warning effectiveness evaluation module; the latter combines historical and real-time data uploaded by stress sensors, past early warning events and their verification results to generate effectiveness evaluation data for the current threshold; this effectiveness evaluation data and construction stage information are input into the reinforcement learning model in the adaptive optimization decision module. The model performs strategy reasoning based on a preset reward function and outputs the threshold adjustment amount for each monitoring area; the early warning threshold dynamic execution module receives the strategy, parses it and sends an update command to the early warning alarm module; the early warning alarm module updates its local threshold table accordingly and continuously compares it with newly collected stress data. Once the limit is exceeded, a safety early warning signal of the corresponding level is triggered; the entire process forms a data-driven closed-loop adjustment mechanism, so that the early warning threshold always adapts to the current structural stress state and construction intention.

[0026] As an optional embodiment, the specific implementation of the present invention is as follows: During the construction of a standard floor in a high-rise residential project, the BIM digital twin model shows that the current stage is the 7th stage of core tube concrete pouring; the construction stage perception module outputs the stage identifier CONCRETE_POURING_L7 accordingly; the early warning effectiveness evaluation module analyzes the stress data of the previous 6 hours and finds that the alarm frequency of pole node R03 under the same load level has increased compared with the previous stage, and most of the alarms are confirmed to be false alarms after on-site verification; the early warning effectiveness evaluation module generates effectiveness evaluation data with a threshold oversensitivity index of 0. 78. It is recommended to lower the threshold of region R03. After receiving the state input, the PPO model in the adaptive optimization decision module infers through the decision network and outputs the adjustment strategy: the threshold of region R03 is lowered, the threshold of region R07 is maintained, and the connection point of horizontal bar H02 is raised. The early warning threshold dynamic execution module parses the strategy into specific values ​​and sends them to the early warning alarm module. After updating the threshold table, the early warning alarm module continues to monitor. In the following 2 hours, no false alarms occurred in region R03, while region H02 successfully captured the actual stress jump and triggered a first-level early warning in a sudden wind-induced vibration coupling loading, verifying the effectiveness of the strategy adjustment.

[0027] Because the construction phase perception module can accurately identify the actual construction conditions of the disc-lock scaffolding, it can provide reliable prior constraints for the early warning strategy, avoiding frequent false alarms caused by low threshold settings during high-load phases. Because the early warning effectiveness evaluation module generates quantitative evaluation data based on historical and real-time stress data and early warning verification results, it can objectively calibrate the actual performance of the current threshold, preventing missed alarms due to biased experience settings. Because the adaptive optimization decision module has a built-in reinforcement learning model and uses construction phase information and effectiveness evaluation data as state inputs, it can autonomously learn the optimal threshold adjustment rules under different combinations of working conditions, improving the system's generalization adaptability to complex construction scenarios. Because the early warning threshold dynamic execution module supports millisecond-level threshold refresh and strategy rollback mechanisms, it can ensure the real-time performance and fault tolerance of the early warning response, avoiding cascading malfunctions caused by incorrect strategies. Because the early warning alarm module performs updated regional threshold comparisons, it can achieve differentiated risk responses in the spatial dimension, improving the overall refined management level of the early warning system.

[0028] As an optional implementation, the construction phase sensing module further includes: The visual progress recognition unit is used to collect image data through cameras deployed at the construction site and to identify the appearance, movement and distribution of key construction components based on computer vision algorithms. The digital twin automatic update unit, connected to the visual progress recognition unit, is used to automatically drive the synchronous update of component status and schedule labels in the BIM digital twin model based on the recognition results. The current construction phase information is automatically determined by the digital twin model based on the updated progress status.

[0029] The visual progress recognition unit can refer to a fixed industrial camera deployed on the end of a tower crane boom, the top platform of a modular scaffold, or the edge protection railing. Its field of view covers the main working surface and typical construction areas of the modular scaffold. The computer vision algorithm used by this visual progress recognition unit is a target detection model based on YOLOv8 or RT-DETR architecture, which is fine-tuned in the modular scaffold construction scenario through transfer learning. It is used to identify the spatial existence, relative displacement trend, and distribution density of key construction components such as formwork, rebar cages, concrete placing booms, and block stacking areas. The recognition results include component category, two-dimensional / three-dimensional spatial coordinates, confidence level, and timestamp. Its output format is a JSON structured data stream, which is transmitted to the digital twin automatic update unit. The digital twin automatic update unit is an embedded edge computing node or a lightweight BIM collaboration platform plugin. It supports IFC standard format parsing and real-time semantic mapping, and can automatically bind the component status output by visual recognition to the attribute fields of the corresponding component objects in the BIM model, and synchronously refresh the schedule tags associated with the component. The BIM digital twin model is a lightweight WebGL model built on the Revit or Bentley platform. It has a built-in schedule logic rule engine, which can automatically deduce and determine the current construction stage based on the component status update sequence and the preset process dependencies. The construction stage information is an enumerated type variable, and the values ​​include, but are not limited to, the foundation support erection stage, the main structure construction stage, the enclosure structure construction stage, the decoration and finishing stage, and the demolition and exit stage. The specific values ​​can be configured according to the actual process flow of the construction site. This embodiment of the invention does not impose any special limitations on this.

[0030] The number of cameras, installation height, tilt angle, and focal length deployed in the visual progress recognition unit can be set according to the lighting conditions, occlusion, and monitoring accuracy requirements of the construction site. The computer vision algorithm used can also be replaced by the Transformer-based ViT-Seg image segmentation model to achieve refined recognition of component boundaries in complex overlapping scenes. The data interaction protocol between the digital twin automatic update unit and the BIM model is WebSocket long connection or OPC UA industrial communication protocol to ensure low-latency synchronization. The semantic definition of the progress plan label adopts the building life cycle stage coding system specified in ISO 15686-7 standard, and a hierarchical label structure can also be customized according to project management specifications. The above implementation methods are all conventional technical paths that can be selected by those skilled in the art according to actual engineering conditions, and the embodiments of the present invention do not impose special limitations on them.

[0031] The visual progress recognition unit continuously collects video streams from the construction site, samples them frame by frame, and inputs them into the target detection model to identify the spatial distribution of the completed installation areas of the formwork supports and the rebar tying points. The recognition results are structured into component IDs, location vectors, and status codes, and pushed to the digital twin automatic update unit. After parsing, the visual progress recognition unit locates the corresponding numbered formwork component family instance in the BIM model, updates its installation status attribute from pending installation to in place, and triggers the event to the progress logic engine. Based on the preset process chain, the engine determines that the completion of formwork installation is a prerequisite for the concrete pouring process, and since there are no other blocking tasks, it automatically updates the overall construction stage to the main structure construction stage—concrete pouring period. This main structure construction stage information is then output as structured parameters to... As an optional implementation method, the early warning effectiveness evaluation module further integrates: The causal inference analysis unit is used to construct a causal graph model based on historical stress data and alarm records, and to quantitatively analyze the strength of the causal relationship between specific construction activities, environmental changes and abnormal stress fluctuations. The Bayesian dynamic evaluation unit receives the output of the causal inference analysis unit and constructs the performance evaluation data into a probability distribution that changes over time. Its prior probability is set based on the historical false alarm / false alarm rate, and its posterior probability is dynamically updated based on the real-time alarm results and on-site verification feedback.

[0032] The causal inference analysis unit is a data-driven attribution module based on a structural causal model (SCM) or Do-calculus framework. Its inputs include multi-source time-series data: historical stress sequences of key nodes of the disc-lock scaffolding, construction logs for the corresponding time periods, environmental monitoring data, and historical early warning trigger records and manual review conclusions. This causal inference analysis unit constructs a directed acyclic graph (DAG) to characterize the potential causal dependencies between variables and uses counterfactual reasoning or propensity score matching methods to estimate the average treatment effect (ATE) of a specific intervention on the target variable, thereby quantifying the strength of the causal relationship. The topology of this causal graph model can be adapted to the actual working conditions of the construction site. For example, in the construction scenario of high-rise buildings, the floor height can be introduced as a hybrid variable node, and in the open-air construction site, the duration of gusts can be added as a moderating variable node. Its specific implementation can be based on the DoWhy library or CausalNex framework in the Python ecosystem, or it can be an embedded lightweight causal inference engine. This embodiment of the invention does not impose any special limitations on this.

[0033] A Bayesian dynamic evaluation unit is a dynamic Bayesian network (DBN) instance that uses performance evaluation data as random variables and time steps as indexes. Its state variable represents the false alarm probability of the current warning threshold in a specific monitoring area. With the probability of missed reports The prior distribution is set based on the historical operational statistics of the early warning and alarm module, for example, using a Beta distribution. ,in and These correspond to the historical cumulative number of confirmed false alarms and the number of times actual risks did not trigger alarms, respectively; the posterior update process follows Bayes' theorem: when a warning is confirmed as a false alarm after on-site verification, the update... When a real risk is confirmed but no alarm is triggered, update. If a risk is confirmed as real and an alarm has been raised, no changes will be made. However, it triggers a confidence weighting of the corresponding path in the causal graph; the performance evaluation data output by the Bayesian dynamic evaluation unit is a set of posterior distribution parameters that evolve over time, which can be used to support the state input of the adaptive optimization decision module; its calculation can be completed locally on the edge gateway, or it can be uploaded to the cloud for collaborative updates through a secure channel, and the specific deployment method can be flexibly configured according to the project's network conditions and data security level.

[0034] The causal inference analysis unit and the Bayesian dynamic evaluation unit form a closed-loop feedback chain: the causal inference results provide structured prior knowledge for the Bayesian unit, and the Bayesian unit then injects the results of each alarm review back into the causal graph model, dynamically correcting the causal strength weights of each side, forming an iterative optimization mechanism of attribution-evaluation-feedback-reattribution; this iterative optimization mechanism does not change the basic comparison logic of the early warning alarm module, but only enhances the interpretability and robustness of its input end—that is, the performance evaluation data on which the early warning threshold generation depends.

[0035] As an optional embodiment, the specific implementation of the present invention is as follows: During the main structure construction phase of a commercial complex project, the system collected data showing early morning stress jumps in the south cantilever area of ​​the 7th floor for three consecutive days; the causal inference analysis unit identified a strong causal correlation between this early morning stress jump and the high-frequency hoisting operations of the tower crane in the early morning, but no significant causal path with structural defects; based on this, the Bayesian dynamic evaluation unit set the prior probability of false alarms for this type of early morning hoisting-induced stress disturbance to 0.65, and in the subsequent 5 similar alarms, 4 were confirmed as non-structural risks after review by the supervisor, and the posterior probability was updated to... The corresponding false alarm probability decreased to an average of 0.6; when the 6th similar alarm was triggered, the system automatically lowered the morning warning threshold for that cantilever area. Sensitivity is assessed, and a briefing is simultaneously sent to management personnel: a stable causal model has been established for the morning hoisting disturbance, and it is recommended to maintain the current threshold observation without emergency response; if a strong gust of wind (wind speed suddenly increases to 12 m / s) is encountered on a certain day, the system detects that the stress response mode deviates from the existing causal path, the Bayesian unit determines the environmental complexity transition, temporarily freezes the posterior update of the cantilever area and triggers the manual review process to ensure that the assessment mechanism does not fail due to abnormal working conditions.

[0036] Through the above technical solutions, this invention achieves the following: By introducing a causal inference analysis unit, the true driving force of construction activities and environmental changes on stress anomalies can be extracted from the apparent correlation, thus improving the ability of early warning effectiveness assessment to identify interference sources; By integrating a Bayesian dynamic assessment unit, the effectiveness assessment data possesses temporal evolution and evidence accumulation, thus enhancing the learning and adaptation ability of the early warning threshold adjustment strategy to on-site feedback; The two work together to solve the technical problem that traditional threshold assessment lacks the ability to deeply attribute alarm causes and cannot distinguish whether stress anomalies are caused by real risks or external interference.

[0037] As an optional implementation, the reinforcement learning model in the adaptive optimization decision module is a multi-agent reinforcement learning architecture, including: The first intelligent agent makes decisions with the goal of maximizing the overall structural safety factor. The second intelligent agent makes decisions with the goal of minimizing interference with construction efficiency. Two agents negotiate and train together using a cooperative game theory framework. The strategy for adjusting the warning threshold is the Pareto optimal solution negotiated by both parties.

[0038] The first intelligent agent is a reinforcement learning agent that is independently modeled in the adaptive optimization decision module and has its own state space and action space. Its reward function takes the overall structural safety factor as the key optimization objective. The overall structural safety factor is dynamically calculated based on finite element simulation results, real-time data from stress sensors, and material performance parameters. The action output of the intelligent agent is the adjustment range of the warning threshold or the level of conservatism enhancement for different monitoring areas. Its strategy tends to improve the warning sensitivity. For example, it automatically lowers the warning threshold in high-risk areas such as densely packed uprights of disc-lock scaffolding and cantilever ends to detect minor stress anomalies in advance. The state input of the intelligent agent includes information on the current construction stage, the historical maximum stress value of each node, the stress change rate of adjacent time periods, and the structural redundancy assessment results. Its specific implementation can be set according to the actual situation. For example, it can be a discrete action intelligent agent based on deep Q-network (DQN) or a continuous control intelligent agent based on proximal policy optimization (PPO). This embodiment of the invention does not make any special limitations on this.

[0039] The second agent is a reinforcement learning agent deployed in parallel with the first agent, sharing some observations but with differentiated reward orientation. Its reward function takes minimizing construction efficiency interference as the core indicator. This core indicator is quantified as the number of process pauses caused by early warning triggering, the response time for manual review, and the probability of hoisting / pouring operation interruption per unit time. The agent's action output is either an upward adjustment of the warning threshold or a tolerance relaxation level. The strategy tends to reduce unnecessary alarms while ensuring basic safety. For example, during the static curing stage after template reinforcement or during routine worker inspections, the stress threshold in non-load-bearing crossbar areas can be appropriately increased to avoid false alarms caused by vibration and noise. The agent's status inputs include construction progress deviation rate, on-site equipment operating status, environmental vibration baseline level, and historical distribution of concentrated false alarm periods. Its implementation can be, for example, a policy gradient agent using an Actor-Critic architecture or a hybrid agent incorporating expert rules. This embodiment of the invention does not impose any special limitations on this.

[0040] The cooperative game framework is a mechanism that supports multi-agent strategy interaction and utility coordination. Embedded in the decision network layer of the adaptive optimization decision module, it receives candidate threshold adjustment schemes generated by the first and second agents and performs joint optimization based on Nash equilibrium constraints and the Pareto front search algorithm. This cooperative game framework constructs a bi-objective optimization problem: ,in Characterizing the security loss function, The efficiency loss function is characterized. Its solution process includes: first, the two agents output initial policy vectors respectively; then, the cooperative game module calculates the weighted Shapley value of their utility functions to allocate negotiation weights; then, the ε-constraint method or weighted sum method is used to iteratively search for the Pareto optimal solution set in the feasible solution space; finally, the solution that minimizes the overall risk entropy is selected as the output policy. The specific implementation of this cooperative game framework can rely on open-source game solving libraries (such as OpenSpiel) or customized multi-objective optimizers. Its parameter configuration, convergence threshold, and weight initialization method can be fine-tuned online according to the actual feedback data from the construction site. This embodiment of the invention does not impose any special limitations on this.

[0041] The Pareto optimal solution is a warning threshold adjustment strategy that, given the current construction stage and real-time stress distribution, cannot further improve one objective without worsening the other. This Pareto optimal solution is manifested as a set of spatially differentiated threshold configurations. For example, a stricter threshold is set for the main beam support point area, a relatively lenient threshold is set for the walkway slab connection node area, and a compromise threshold is used for the diagonal brace intersection node. The generation process of this Pareto optimal solution does not rely on preset rules but is completed through multiple rounds of strategy trial and error and utility feedback between two agents in a cooperative game framework. Its stability can be verified through historical working conditions. For example, in simulations during peak concrete pouring periods, this Pareto optimal solution can increase the alarm trigger rate in the core support area while decreasing the false alarm rate in non-critical areas.

[0042] The working process of the multi-agent reinforcement learning architecture is as follows: When the adaptive optimization decision module receives the mid-stage marker of concrete pouring output by the construction stage perception module and the current false alarm rate and false alarm rate performance evaluation data output by the early warning performance evaluation module, the first agent calculates based on the structural safety factor model that the early warning threshold in the A1-A5 pole group area needs to be lowered, and the second agent calculates based on the construction efficiency interference model that the threshold in the B1-B3 walkway area can be raised. The cooperative game framework receives the proposals from both agents and, combined with auxiliary variables such as the current wind speed, temperature and humidity, and the on-duty status of the supervisors, constructs a dual-objective optimization model. After 5 rounds of iteration, it outputs the Pareto optimal strategy: lower the threshold in A1-A5, raise the threshold in B1-B3, and keep the baseline threshold unchanged in the C1-C4 diagonal brace area. This Pareto optimal strategy is then parsed by the early warning threshold dynamic execution module and sent to the early warning alarm module.

[0043] As an optional embodiment, the specific implementation of the present invention is as follows: During the construction of a standard floor in a high-rise residential project, the system identifies that it is currently in the shear wall concrete pouring stage, and the visual progress recognition unit confirms that the concrete placing boom is in place and the pump pipe is connected; the digital twin automatic update unit synchronously updates the load condition label of the corresponding area in the BIM model to be dominated by dynamic impact load; the early warning effectiveness evaluation module determines that the current structure is in the critical zone of elastic deformation based on the stress sequence analysis of the past 3 hours; At this point, the adaptive optimization decision-making module initiates multi-agent collaborative reasoning: the first agent, based on the finite element pre-simulation results, suggests lowering the warning threshold for the core shear wall support pole group from 120MPa to 95MPa; the second agent, based on the tower crane scheduling plan and vibration operation rhythm, suggests raising the threshold for the operating platform connecting crossbar group from 85MPa to 92MPa; the cooperative game framework integrates contextual information such as the on-site wind speed of 4.2m / s, the concrete placement temperature of 28℃, and the online confirmation status by the supervising engineer to generate a Pareto index. The optimal solution is to set the thresholds for S1-S8 to 96.5 MPa and the thresholds for H1-H6 to 91.8 MPa, with the additional condition that if the stress rise slope is greater than 0.8 MPa / s within the next 15 minutes, the emergency threshold reduction mode for S1-S8 will be automatically activated. This Pareto optimal strategy was verified by the dynamic threshold simulation unit to be effective without alarm storm risk. Ultimately, it successfully avoided two potential false alarms caused by pump pressure fluctuations during the peak pouring period, and captured the precursor stress distortion of local buckling of the S3 pole 17 seconds in advance.

[0044] Through the above technical solution, this invention achieves the following: setting up a first agent and a second agent with the objectives of maximizing the overall structural safety factor and minimizing construction efficiency interference, respectively, which can simultaneously consider safety and construction continuity during the adjustment of the early warning threshold; the two agents negotiate strategies through a cooperative game framework and output Pareto optimal solutions, avoiding excessive conservatism in safety or blind sacrifice of efficiency caused by single-objective optimization; the Pareto optimal solution is reflected in spatially differentiated threshold configuration and supports dynamic context response, which can adapt to the differences in load distribution and risk characteristics of disc-lock scaffolding under different construction stages such as concrete pouring, masonry, and formwork removal, thereby improving the engineering applicability and on-site acceptability of the early warning system.

[0045] As an optional embodiment, the federated learning aggregation unit is used to periodically receive locally optimized threshold adjustment model parameters from cloud edge gateways of multiple similar projects, while ensuring the data privacy of each construction site, and to perform secure aggregation to generate a globally optimized model. The dynamic threshold simulation and deduction unit, based on the aggregated model and the current digital twin of the disc-lock scaffolding, performs short-term stress response simulation on the threshold to be adjusted, predicts the alarm modes that may be triggered after the adjustment, and if the simulation shows that it may cause an alarm storm or cover up the real risk, it triggers the adjustment strategy rollback and manual review request.

[0046] The federated learning aggregation unit is deployed on the central cloud platform and is a distributed model fusion component. Its input consists of local reinforcement learning model parameter updates uploaded from edge computing nodes at multiple similar disc-lock scaffolding construction sites, excluding raw stress data, image data, or environmental sensor data. This federated learning aggregation unit employs a weighted average or differential privacy-preserving secure aggregation algorithm, for example, satisfying... A noise injection aggregation mechanism with differential privacy constraints is used to ensure that the data of each participant is irreversible. The global optimization model is a shared policy network weight under a multi-task learning framework. Its output space covers the threshold offset suggestions of each monitoring area under different construction stages. The model structure can be set according to the actual situation. For example, a lightweight Transformer encoder or a graph neural network (GNN) can be used to model the structural topology differences across construction sites. This embodiment of the invention does not impose any special limitations on this.

[0047] The dynamic threshold simulation and derivation unit is integrated into the digital twin platform and is a finite element simulation engine module. Its inputs include: candidate threshold adjustment schemes output from the global optimization model generated through federated aggregation, the geometric configuration, material parameters, boundary constraints, and real-time loading conditions of the current digital twin of the disc-lock scaffolding. This dynamic threshold simulation and derivation unit, by calling an open-source solver, completes short-term (≤30 minutes) stress response time history simulations under the corresponding adjustment strategy within seconds, outputting the stress envelope of key nodes, alarm trigger frequency distribution, and multi-region linkage alarm probability matrix. An alarm storm is defined as... Within 5 sampling periods, ≥3 warnings are triggered in the same monitoring area, or warnings are triggered concurrently across ≥5 adjacent pole nodes within a 10-second window; masking the true risk is defined as the simulation showing that the stress peak corresponding to a certain type of typical over-limit working condition has exceeded the design safety factor threshold, but no warning is triggered due to the threshold being adjusted upwards; when any of the above situations are identified, the dynamic threshold simulation and deduction unit automatically returns a strategy rejection signal to the adaptive optimization decision module, and simultaneously pushes a manual review request instruction to the on-site management personnel terminal, including the threshold adjustment items to be reviewed, the simulation risk heat map, and the index of historical similar decision cases.

[0048] The collaborative working process between the federated learning aggregation unit and the dynamic threshold simulation and deduction unit is as follows: After a construction site edge gateway completes a round of local reinforcement learning model training, it encrypts and uploads its gradient update to the central cloud platform; the federated learning aggregation unit completes parameter aggregation under the condition of satisfying the preset privacy budget, and generates an updated global policy model; this reinforcement learning model, along with the current digital twin state, is pushed to the dynamic threshold simulation and deduction unit; the latter constructs simulation boundary conditions based on the finite element mesh driven by the current disc-lock scaffolding BIM model, the measured material degradation coefficient, and the meteorological load spectrum of the past 24 hours, and conducts Monte Carlo sampling simulation on the threshold adjustment vector to be issued to evaluate its alarm stability under various potential disturbance conditions; only when the simulation confirms that the alarm mode is within the controllable range is the global policy allowed to be written into the runtime cache of the early warning threshold dynamic execution module and take effect; otherwise, the rollback process is initiated to maintain the previously verified threshold configuration.

[0049] As an optional embodiment, the specific implementation of the present invention is as follows: Taking a high-rise residential project A as an example, it is in the concrete pouring stage. The stress sensor detects that the stress of the uprights in the core tube area is continuously rising but has not yet reached the original fixed threshold. At this time, the local reinforcement learning model suggests temporarily increasing the threshold of the core tube area by 15%. After the parameter update is uploaded to the federated learning aggregation unit, it combines the parameter updates from 8 other similar projects under construction across the country to generate a global optimization model. The new threshold suggestion output by the global optimization model is sent to the dynamic threshold simulation and deduction unit. The simulation results show that if this scheme is implemented, when a sudden gust of wind is encountered in the following 2 hours, coupled with the vibration excitation of the pump truck, it will cause 6 adjacent monitoring points to trigger warnings continuously within 30 seconds, forming an alarm storm. The system triggers a rollback accordingly and pushes an AR visualization briefing to the project director's terminal, presenting the wind-vibration coupling path, the topological relationship of high-risk nodes, and alternative threshold suggestions in the simulation for human decision-making reference.

[0050] Through the above technical solutions, this invention achieves the following: By introducing federated learning aggregation units, secure sharing of threshold optimization experience across projects is realized without exposing the original stress data of each construction site, thus improving the generalization and adaptability of the global model to new construction techniques or regional environmental disturbances; By setting up dynamic threshold simulation and deduction units, short-time stress response simulation verification is carried out based on the current digital twin before each threshold adjustment, thus avoiding the false alarm propagation or risk masking problems caused by unverified strategies directly acting on the physical system, and enhancing the robustness and reliability of the early warning system.

[0051] As an optional implementation, the multimodal fusion alarm unit simultaneously receives data streams from stress sensors, vibration / tilt data streams from environmental sensors, and real-time image streams from the visual progress recognition unit in Embodiment 2 at its input terminal. An attention-based alarm decision engine is used to perform weighted fusion analysis on the above multimodal data streams. The highest level of deterministic alarm is triggered only when evidence from multiple modalities points to the same high-risk event. If only a single modality is abnormal, a lower level of suggestive check alarm is triggered.

[0052] The multimodal fusion alarm unit is integrated into the data access and preprocessing module of the edge computing gateway or cloud early warning server. It is configured to support concurrent access of multiple protocols and can simultaneously receive and cache heterogeneous time-series data and unstructured image frames from different sensing channels. The data stream of the stress sensor is the raw voltage signal of axial / shear stress acquired at a sampling rate of 10 Hz. After conversion by the built-in ADC and temperature compensation, it is output as a standardized stress value sequence. The vibration / tilt data stream from the environmental sensors is a triaxial acceleration and biaxial tilt analog output, which is low-pass filtered and attitude calculated to generate a structural dynamic response feature vector. The real-time image stream from the visual progress recognition unit is an H.264 encoded video stream with a resolution of 1920×1080 and a frame rate of 5 fps, or keyframe images captured as needed. Its content covers the connection nodes of the uprights of the disc-lock scaffolding, the deformation area of ​​the crossbars, and the activity range of personnel on the working surface. Each data stream carries a unified timestamp and spatial coordinate identifier when it is accessed, so as to facilitate subsequent cross-modal alignment. The hardware implementation of the multimodal fusion alarm unit can be set according to the actual situation. For example, it can be an embedded ARM Cortex-A72 platform or a Docker containerized service deployed on an industrial gateway. This embodiment of the invention does not impose any special limitations on this.

[0053] The attention-based alarm decision engine is a lightweight neural network model deployed locally or at the edge of the early warning alarm module. Its structure includes: a multi-source feature encoding sub-network, a cross-modal attention weight generation sub-network, and a collaborative evidence discrimination sub-network. The multi-source feature encoding sub-network extracts features from three types of input data respectively: one-dimensional convolution + BiLSTM encoding is used for stress data stream to generate time-sensitive stress state representation; wavelet packet decomposition + fully connected mapping is used for vibration / tilt data stream to extract frequency domain dominant modal response features. The MobileNetV3 backbone network is used to extract spatial local anomaly hotspot features from the image stream, and ROIAlign is used to focus on the button node region. The cross-modal attention weight generation subnetwork is based on the self-attention mechanism, using each modal feature as the input Query, Key and Value, dynamically calculating the semantic correlation matrix between the three, and outputting a normalized attention weight vector. The collaborative evidence discrimination subnetwork inputs the weighted fused joint features to the binary classification fully connected layer and outputs the decision result of deterministic alarm or suggestive inspection alarm. The model parameters of the alarm decision engine can be quantized, compressed and pruned according to the actual deployment conditions of the construction site. For example, it can be deployed as an INT8 precision TensorRT inference engine, or it can be replaced with a lightweight MLP structure obtained by knowledge distillation. This embodiment of the invention does not make any special limitations on this.

[0054] The alarm decision engine based on the attention mechanism first synchronizes and encodes the received stress data stream segments, vibration / tilt angle data stream segments, and image frames during operation. Then, the consistency of the three is evaluated through an attention weight generation sub-network: when the stress value is in the critical range, the tilt angle change rate exceeds 0.1° / s, and the corresponding node area in the image shows pixel-level displacement, the attention mechanism automatically assigns high weights to the three, the fused features deviate from the normal distribution, and triggers the highest level of deterministic alarm. When there is only a brief spike in stress data, and the vibration and image do not have abnormal responses, the attention mechanism identifies it as transient interference, reduces the weight of the stress channel, the fused features do not reach the alarm threshold, and only outputs a low-level prompt inspection alarm, along with the sensor number and time window for suggested review. This operation does not rely on manually set rules, but is entirely driven by the trained attention weights, and its judgment logic can be continuously fine-tuned online with the feedback data from the field.

[0055] As an optional embodiment, the specific implementation of the present invention is as follows: During the concrete pouring operation of a standard floor in a high-rise residential project, a stress sensor detected a continuously rising axial stress at the core support column, reaching 82% of the design bearing capacity; simultaneously, a MEMS tilt sensor installed on the top of the same column recorded a slow tilting trend of 0.15°; at the same time, the visual progress recognition unit captured a slight warping deformation of the crossbar connecting plate near the core support column, and the image feature encoding output deviated from the historical template; after attention-weighted fusion, the alarm decision engine determined that the three types of evidence were highly coordinated, output a structural instability precursor label, and triggered a red audible and visual alarm and a bright flashing of the corresponding component in the BIM digital twin model; in the subsequent demolding and cleaning stage, a sensor generated a short-term noise pulse due to dust obstruction, and the stress data stream showed an isolated spike, but the tilt angle and image stream remained stable. The attention mechanism automatically suppressed the influence of this channel, and the fusion result did not reach the alarm threshold. The system only pushed a suggestive inspection alarm to the mobile terminal, recommending cleaning the #A3-07 stress sensor lens.

[0056] Through the above technical solutions, this invention achieves the following: Because the multimodal fusion alarm unit simultaneously accesses three heterogeneous data streams—stress, environmental vibration / tilt, and visual images—it can overcome the information limitations of a single sensing dimension; because the alarm decision engine based on the attention mechanism performs weighted fusion based on the spatiotemporal consistency of multi-source evidence and sets collaborative pointing to the same high-risk event as a prerequisite for high-level alarms, it avoids false alarms caused by single-point disturbances; because it only triggers a suggestive inspection alarm rather than an emergency response for single-modal anomalies, it reduces alarm fatigue among on-site personnel and improves the effectiveness and rationality of safety response levels.

[0057] As an optional implementation method, system integration: The swarm intelligence optimizer uses the ant colony algorithm to simulate multiple virtual agents exploring the policy space in parallel. Each agent uses the initial negotiation results of the cooperative game framework as the initial pheromone distribution to perform a local fine-grained search. The dynamic environment adaptability assessment unit constructs a dynamic environment complexity index based on real-time meteorological data, equipment operating status, and worker behavior pattern data. When the index exceeds the threshold, it automatically enhances the exploration weight of the swarm intelligence optimizer to discover emergency strategies to adapt to sudden and complex situations.

[0058] The swarm intelligence optimizer is a collaborative search mechanism embedded within a multi-agent reinforcement learning architecture. Its core function is to further explore the policy space with fine-grained precision after initially achieving a Pareto optimal solution in cooperative game theory. This swarm intelligence optimizer is a distributed search module implemented based on the ant colony algorithm (ACO). Each virtual agent represents a potential combination of warning threshold adjustment strategies. When the agent moves in the policy space, it selects a path based on the pheromone concentration. The initial pheromone distribution is generated by mapping the preliminary negotiation results output by the first agent and the second agent under the cooperative game framework. The pheromone update rule can be dynamically adjusted based on the reward feedback obtained by each agent in the simulation in the digital twin environment. For example, if the threshold configuration triggered by a certain agent reduces the false alarm rate and does not cause a false alarm in the simulation, the pheromone concentration of the corresponding path will be enhanced. The parameter configuration of this swarm intelligent optimizer includes the number of agents, the pheromone evaporation coefficient, and the heuristic factor weight, where the number of agents is 50 to 200 or 300. The ant colony algorithm can be implemented in a discrete or continuous manner and is suitable for joint optimization of multi-dimensional early warning threshold vectors, which include stress thresholds, vibration thresholds, and tilt angle thresholds corresponding to different monitoring areas.

[0059] The dynamic environment adaptability assessment unit is a functional module used to quantify the real-time disturbance intensity at the construction site. Its input data includes wind speed, rainfall intensity, and temperature gradient data collected by meteorological sensors; start-up and shutdown status and vibration spectrum characteristics of large equipment such as tower cranes / pump trucks reported by the edge gateway; and worker density, movement trajectory abrupt change rate, and work clustering data obtained through video analysis or UWB positioning systems. This dynamic environment adaptability assessment unit maps the above heterogeneous data into a single scalar—the dynamic environment complexity index—through a preset weighted fusion model. The calculation formula is expressed as: These are adjustable weighting coefficients, summing to 1, and pre-configured according to the project type. These are the normalized meteorological disturbance component, equipment disturbance component, and personnel behavior disturbance component, respectively; the value range of this dynamic environmental complexity index is [0,1]. When the environment is determined to be highly complex, the dynamic environment adaptability assessment unit outputs a control signal to the swarm intelligence optimizer, dynamically increasing its exploration weight from the base value of 0.3 to the enhanced value of 0.6 to 0.8, thereby expanding the search range of the ant colony agent in the policy space and increasing the sampling probability of unconventional threshold combinations; the output response delay of the dynamic environment adaptability assessment unit is less than 500ms or 200ms.

[0060] The collaborative working process between the swarm intelligence optimizer and the dynamic environment adaptability assessment unit is as follows: Under normal construction conditions, when the basic complexity index output by the dynamic environment adaptability assessment unit is below the threshold, the swarm intelligence optimizer maintains the default exploration weights and performs a fine search within the neighborhood of the initial solution in the cooperative game, gradually fine-tuning the warning thresholds of each region to approach a better Pareto front; when a typhoon warning is issued, a sudden downpour occurs on-site, or large-scale cross-operations take place, the meteorological and behavioral data change drastically, leading to... As the threshold rapidly rises and exceeds the limit, the evaluation unit immediately triggers an instruction to enhance the exploration weight. The ant colony agent then expands its step size, increases the probability of random jumps, and launches a new round of parallel exploration in a wider policy space, thereby quickly identifying a new threshold configuration scheme that is suitable for extreme conditions. This process does not change the structural constraints of the original cooperative game framework, but only dynamically adjusts the intensity of the search behavior to ensure that the policy update always meets the dual-objective equilibrium boundary of safety and efficiency.

[0061] As an optional embodiment, the specific implementation of the present invention is as follows: During the concrete pouring stage of a standard floor of a super high-rise building, the system initially operates in conventional mode. The swarm intelligence optimizer performs local optimization around the initial solution of the cooperative game with an exploration weight of 0.35. The initial solution of the cooperative game sets the threshold for the critical pole area to 120MPa and the threshold for the horizontal scissor bracing area to 85MPa. When the meteorological platform pushes a warning that the gusts will reach level 10 in the next 2 hours, and at the same time, video analysis identifies that 3 tower cranes are rotating at high frequency simultaneously and the worker density on the work surface has increased to 1.8 times the threshold, the dynamic environmental adaptability assessment unit calculates in real time... Then, a weight enhancement signal is sent to the swarm intelligence optimizer; The optimizer increased the exploration weight to 0.72, and the 50 virtual agents immediately jumped out of the original strategy neighborhood. Among them, 12 agents tried to temporarily increase the pole area threshold to 135-142 MPa to tolerate wind-induced additional stress, 8 agents explored the introduction of a dynamic attenuation coefficient in the scissor bracing area, with the dynamic attenuation coefficient scaling the threshold linearly according to wind speed, and another 5 agents tested the activation of a graded alarm mechanism during periods of strong disturbance. After verification by digital twin simulation, the combination of pole area threshold + 12% + 0.85 times dynamic attenuation in the scissor bracing area proposed by agent A17 reduced the number of false alarms from 4.2 times per hour to 0.9 times while maintaining a zero false alarm rate. This strategy was adopted and sent to the early warning threshold dynamic execution module for real-time updates.

[0062] Through the above technical solutions, this invention achieves the following: By setting up a swarm intelligence optimizer and using the ant colony algorithm to conduct parallel and refined searches based on the initial solution of cooperative game theory, the local optimization capability of multi-agent reinforcement learning in the policy space is improved; By setting up a dynamic environment adaptability assessment unit and constructing an environmental complexity index based on meteorological, equipment, and personnel behavior data and enhancing the exploration weight when limits are exceeded, the system can proactively expand the search range under sudden and complex working conditions such as typhoons, rainstorms, and dense cross-operations, breaking through the inertia of the original strategy and quickly generating early warning threshold adjustment strategies adapted to emergency scenarios; Since both the swarm intelligence optimizer and the dynamic environment adaptability assessment unit operate within the multi-agent reinforcement learning architecture, and all input data comes from sensors already deployed at the construction site and existing business systems, no new hardware entities or communication links are introduced, making it fully compatible with the existing modular scaffolding safety early warning system architecture.

[0063] As an optional implementation, the social sentiment computing module uses sentiment computing algorithms to analyze the decision preference patterns of construction site managers in the trade-off between safety and efficiency, based on their historical decision data. The strategy explanation and visualization unit transforms the decision-making logic, trade-offs, and rationality of the final strategy of each agent during the negotiation process into an understandable security briefing using natural language generation technology, and then visualizes it using augmented reality devices to help managers understand and trust the system's decision-making.

[0064] The social sentiment computing module is a software functional module deployed on an edge server or cloud platform. The input data includes historical early warning response records, on-site verification conclusions, operation logs of management personnel during the handling of alarm events, as well as associated construction stage tags and environmental context information. The sentiment computing algorithm used by this social sentiment computing module is a sentiment tendency modeling model based on LSTM or Transformer architecture, which is used to extract implicit preference dimensions such as safety conservatism, efficiency sensitivity, and risk tolerance from unstructured text and time-series operation behaviors, and quantify them into a continuous preference vector. The value range and update cycle of this continuous preference vector are set according to the actual situation. For example, the model parameters are retrained once every quarter based on new decision samples. This embodiment of the invention does not make any special limitations on this.

[0065] The strategy interpretation and visualization unit is integrated into a hardware-software co-processing module in the system management terminal or mobile inspection equipment. Its natural language generation technology is a lightweight, finely tuned T5 or BART model. The input consists of the state-action pairs, reward allocation weights, Pareto front search paths, and the constraint satisfaction of the final policy for each agent during the multi-agent reinforcement learning process. The output is a structured safety briefing, which includes, but is not limited to: the first agent suggests reducing the threshold of area A by 12% because its current stress growth slope is close to the critical value before the historical collapse; the second agent opposes the adjustment because it is expected to affect the subsequent rebar hoisting rhythm. Through collaborative negotiation, a compromise solution was adopted: the threshold for Zone A was lowered by 8%, and a temporary reinforcement reminder for Zone B was simultaneously activated. This structured safety briefing can be further mapped to the spatial coordinate system of the physical scaffolding identified by augmented reality devices. Floating labels, heat arrows, and causal relationship maps are overlaid on the corresponding uprights / horizontal bars using AR glasses, achieving spatial anchoring and layered deployment of decision-making basis. The augmented reality device is HoloLens 2, Rokid Max, or an industrial-grade AR terminal supporting SLAM positioning. The displayed content can dynamically switch abstract levels according to the manager's gaze focus. This embodiment of the invention does not limit the specific model, optical scheme, or interaction method.

[0066] The social sentiment computing module collaborates with the policy interpretation and visualization unit: After the multi-agent reinforcement learning model completes a round of policy negotiation, the manager preference vector output by the social sentiment computing module is injected into the prompt engineering template of the policy interpretation and visualization unit. This guides the natural language generation model to prioritize explanatory elements that align with the manager's cognitive habits—for example, for a manager with a preference for safety, emphasis is placed on critical state identification and redundancy margin retention; for a manager with a preference for efficiency, the focus is on explaining construction interruption time estimates and process connection guarantee measures. This collaborative mechanism does not change the output of the reinforcement learning model itself, but only affects its external expression and presentation, thereby improving the human-machine collaboration adaptability without sacrificing the scientific nature of decision-making.

[0067] As an optional embodiment, the specific implementation of the present invention is as follows: During the main structure construction phase of a high-rise residential project, historical data from the project manager showed that in nearly 30 early warning events, he chose to immediately suspend work and organize a comprehensive investigation in 26 of them, reflecting a conservative tendency towards safety; the social sentiment calculation module generates a preference vector [0.92, [0.21]; When the adaptive optimization decision module generates a Pareto optimal solution with a 10% reduction in the threshold of area A based on the current concrete pouring conditions and real-time stress trends, the strategy explanation and visualization unit automatically constructs a briefing: the stress increase in area A reaches 0.83 MPa / min, approaching the historical threshold of 0.85 MPa / min before collapse, and lowering the threshold can capture potential instability 12 minutes in advance; the threshold in area B is simultaneously raised to maintain the overall construction rhythm, and a red pulse light effect and critical warning: +12min buoy are projected on the top of the pole in area A through AR glasses, and a green progress bar and rhythm guarantee: +3.2h mark are displayed on the side wall of the crossbar in area B; the on-site management personnel intuitively understand the adjustment motivation by staring at the buoy to trigger voice broadcast and overlay of three-dimensional stress cloud map, without making a request for manual intervention, and the system strategy is executed smoothly.

[0068] Through the above technical solutions, this invention achieves the following: by introducing a social sentiment computing module to model the historical decision-making data of management personnel, the cooperative game framework can proactively adapt to human cognitive habits during the strategy negotiation process; by converting the multi-agent decision-making logic into natural language briefings and combining them with augmented reality devices for spatial presentation, the semantic gap between human and machine decision-making can be reduced, and the depth of understanding and trust of on-site management personnel in the system output can be improved; thus, without changing the essence of the early warning threshold adjustment, the acceptance resistance brought about by highly automated decision-making can be effectively alleviated, and the sustainable application of intelligent safety early warning systems in real construction scenarios can be promoted.

[0069] As an optional implementation, the federated learning aggregation unit also includes: The anomaly model detection and defense module uses adversarial machine learning technology to perform feature analysis on the local models submitted by each construction site before parameter aggregation, and to detect and isolate malicious or low-quality model updates. The adaptive weighted aggregation optimizer dynamically adjusts the weight of each construction site in the federated aggregation based on its data quality score, model update stability index, and historical contribution. The data quality score is calculated by combining the output of the anomaly model detection and defense module, the construction site supervision score, and the historical early warning accuracy.

[0070] The anomaly detection and defense module is deployed on the cloud edge gateway or central server side. It is a lightweight model verification unit used to perform multi-dimensional feature analysis on the parameter distribution, gradient direction, loss function sensitivity, and robustness to adversarial examples after receiving locally optimized and threshold-adjusted model parameters uploaded from various construction sites. The anomaly detection and defense module is based on an anomaly detection model with a generative adversarial network (GAN) discriminator structure or an isolated forest. Its input includes the statistical moment features of the local model weight matrix, the L2 norm sequence of the gradient update trajectory during training, and the prediction confidence fluctuation curve on several standard stress perturbation samples. The anomaly detection and defense module does not change the original structure of the local model, but only performs non-intrusive evaluation. When it detects that the parameters deviate from the normal cluster center by more than a preset threshold, or show abnormally high sensitivity to minor perturbations, it is judged as malicious or low-quality update and logically isolated from the current aggregation queue, and does not participate in the subsequent weighted aggregation process. The detection rules of the anomaly detection and defense module dynamically update the baseline threshold through a sliding window method according to the data noise level under different construction environments. In the rainy season high humidity and vibration environment, the gradient consistency tolerance is relaxed to 1.3 times the original set value to avoid misjudging reasonable model offsets caused by sensor drift as anomalies.

[0071] The adaptive weighted aggregation optimizer, integrated into the federated learning coordination service, serves as the weight scheduling unit. Its output aggregate weights for each site are normalized real values, with a range of [value missing]. Furthermore, the sum of the weights of all participating construction sites is 1; this adaptive weighted aggregation optimizer is a weighted scoring model based on the fusion of multi-source indicators, where the data quality score is composed of three weighted components: the credibility score output by the anomaly model detection and defense module (weight 0.4), the standardization score of construction data regularly filled in and encrypted and uploaded by on-site supervisors (weight 0.3), and the percentage conversion value corresponding to the construction site's early warning accuracy rate in the past 30 days (defined as the proportion of real risk events successfully captured) (weight 0.3). The model update stability index is the standard deviation of the model parameter change amplitude in five consecutive local training sessions, used to characterize the convergence status of the local model. If the standard deviation is lower than 0.008, it is considered that the update is too smooth, and its weight is reduced accordingly. The historical contribution is the ratio of the number of updates of the construction site that have been adopted into the global model in the past 6 months to the total number of submissions, used to reflect its long-term collaborative reliability. All the above indicators are updated monthly and uploaded after adding Laplace noise through a differential privacy mechanism. The noise scale parameter ε=1.2 ensures that individual construction site data is not traceable.

[0072] The anomaly detection and defense module works in conjunction with the adaptive weighted aggregation optimizer: at the beginning of each federated aggregation cycle, the anomaly detection and defense module first performs a preliminary screening of the credibility of all models to be aggregated, outputting a preliminary set of usable construction sites and their corresponding credibility labels; then, the adaptive weighted aggregation optimizer calculates the aggregation weight of each construction site only within this preliminary set of usable construction sites, based on real-time updated multidimensional indicators; after the weight calculation is completed, the parameter vectors of the retained models are aggregated using a weighted average method to generate the global optimization model for this round; if a construction site is judged as anomaly twice in a row, a manual review process is automatically triggered, and its upload permission for the next round is suspended until the project management confirms that the data quality has been restored before it can be reconnected; this differential privacy mechanism not only ensures the security boundary of the federated learning process, but also maintains the system's adaptability and resilience in heterogeneous construction site environments.

[0073] As an optional embodiment, the specific implementation of the present invention is as follows: In a cross-site federated learning scenario involving a high-rise residential project A and an industrial plant project B, project A recently experienced zero-point drift of its stress sensor due to tower crane vibration. Its local model exhibited an abnormally high false alarm tendency in adversarial sample testing. The abnormal model detection and defense module, by analyzing its gradient direction dispersion (2.8 times the standard deviation higher than the historical mean) and disturbance robustness decay rate (a decrease of 47%), determined it to be a low-quality update and isolated it. Meanwhile, project B's early warning accuracy rate reached 92.6% over the past three months, the supervisor's score remained stable at 96 points, and the model update stability index was better than the average level. Based on this, the adaptive weighted aggregation optimizer increased its weight from the baseline of 0.35 to 0.48. Finally, in the subsequent 72-hour stress test, the global model obtained by aggregation shortened the response time for identifying stress mutations in the key upright area during the concrete pouring stage by 1.3 seconds, reduced the false alarm rate, and did not exhibit an alarm storm phenomenon, verifying the actual improvement effect of the dual protection mechanism on the quality of federated learning.

[0074] Through the above technical solutions, this invention achieves the following: Setting up an anomaly model detection and defense module to perform pre-verification of the local model based on adversarial machine learning technology, effectively blocking malicious or inaccurate models from polluting the global aggregation results; configuring an adaptive weighted aggregation optimizer and introducing three dynamic indicators: data quality score, update stability, and historical contribution, to achieve differentiated weight allocation while ensuring data privacy, enabling high-quality data sources to have a positive impact on the evolution of the global model; and integrating three dimensions—technical (anomaly detection output), management (supervision score), and business (early warning accuracy)—to comprehensively reflect the actual operational level of the construction site, promoting a positive closed loop where high-quality data drives high-quality models, and high-quality models in turn support accurate early warnings, thereby improving the robustness, fairness, and engineering applicability of the federated learning mechanism in multi-construction site collaborative early warning scenarios.

[0075] As an optional implementation, the dynamic threshold simulation and deduction unit also includes: The virtual-real linkage verification mechanism applies micro-excitation through programmable micro-vibrators installed at key nodes of the physical disc-lock scaffolding while simultaneously performing digital twin simulation, and measures the actual dynamic response. The model corrector based on transfer learning takes the difference between virtual and real response data as input, uses a deep residual network to learn the systematic error characteristics between the digital twin model and the actual structure, and corrects the finite element simulation parameters in real time, thereby improving the consistency between the prediction results and the physical reality.

[0076] The virtual-real linkage verification mechanism simultaneously activates programmable micro-exciters deployed at the connection nodes of the uprights, horizontals, or diagonal braces of the physical disc-lock scaffold during stress response simulation and deduction in the digital twin model. These programmable micro-exciters can output micro-amplitude sinusoidal frequency sweep or random excitation signals with a frequency range of 1 Hz to 200 Hz and an excitation amplitude controllable within ±0.05 mm. The micro-amplitude excitation does not change the normal load-bearing state of the disc-lock scaffold, but is sufficient to excite its first to third-order modal vibration response. The measured actual dynamic response data includes, but is not limited to, acceleration time history curves, frequency response functions (FRF), mode shapes, and natural frequencies. These data are acquired by triaxial accelerometers arranged at the same node and uploaded in real time to the edge computing node through an industrial IoT gateway. The data is then aligned with the simulation output of the digital twin model at the millisecond level and compared with its features.

[0077] The programmable micro exciter is an electromagnetic or piezoelectric ceramic micro actuator with dimensions no larger than Φ25 mm × 40 mm. Installation methods include magnetic base fixing, threaded embedded mounting, or clamp-type quick-release structure. It is compatible with the outer diameter of common Q345 steel pipes and end flange interfaces of disc-lock scaffolding. Its control signal is generated by an edge-side FPGA module, supporting remote configuration of excitation waveform, amplitude, and duration. During non-simulation periods, the programmable micro exciter is in a dormant state with power consumption below 1W, ensuring continuous operation of daily monitoring tasks for disc-lock scaffolding.

[0078] The transfer learning-based model corrector uses the difference between simulated and measured natural frequencies—including the deviation between simulated and measured natural frequencies, mode shape MAC values, and frequency response function amplitude error matrices—as a supervisory signal to construct an input feature tensor. This feature tensor, after preprocessing, is input into a lightweight deep residual network. The network output is an incremental vector of key parameters to be corrected in the finite element model, including the steel pipe elastic modulus correction coefficient, node connection stiffness attenuation factor, and foundation constraint spring stiffness offset. The deep residual network is deployed on an edge server with an inference latency of less than 500 ms, supporting online parameter updates once per simulation cycle. The model training data comes from a historical set of virtual-to-real comparison samples under various working conditions and has been adapted across construction sites before deployment, enabling generalization to different specifications of modular scaffolding and foundation conditions.

[0079] The difference between virtual and real response data is measured in a joint frequency and time domain. For example, the relative error of each modal frequency is calculated in the frequency domain, and the root mean square error (RMSE) and dynamic time warping (DTW) distance of the acceleration response signal are calculated in the time domain. The difference features are normalized to form the input of the model corrector, avoiding gradient imbalance caused by dimensional differences. The extraction and encapsulation of difference features do not depend on manually set thresholds, but are automatically completed by the embedded signal processing module.

[0080] The finite element simulation parameters include material constitutive parameters, contact interface friction coefficient, boundary condition constraint stiffness, and initial geometric defect amplitude of the component; the elastic modulus correction coefficient ranges from 0.85 to 1.15, the node connection stiffness attenuation factor ranges from 0.3 to 1.0, and the foundation constraint spring stiffness offset can be mapped in stages according to the characteristic value of foundation bearing capacity in the geological survey report; all parameter corrections are injected in real time in the local finite element solver of the digital twin without restarting the simulation process.

[0081] The virtual-real linkage verification mechanism and the transfer learning-based model corrector work together to form a closed-loop calibration process of simulation-excitation-acquisition-comparison-inversion-correction-resimulation: Before each early warning threshold adjustment strategy is generated, the system first calls the digital twin to perform a finite element transient response simulation under the target working condition; then it triggers the micro exciter to apply the preset excitation and simultaneously acquires the measured response; then it extracts the virtual-real difference features and sends them to the model corrector to obtain the parameter correction amount; finally, it injects the corrected parameters into the simulation model and reruns the simulation to output a high-fidelity stress response prediction, supporting the reliability assessment of the subsequent threshold adjustment strategy.

[0082] Through the above technical solutions, this invention achieves the following: By setting up a virtual-real linkage verification mechanism, the real dynamic response of the physical disc-lock scaffold can be obtained synchronously during the digital twin simulation, overcoming the systematic deviation caused by the simplification of modeling in pure numerical simulation; By introducing a model corrector based on transfer learning and driving the real-time correction of finite element parameters by the difference between virtual and real responses, the influence of implicit degradation factors such as material performance dispersion, node loosening, and foundation settlement can be dynamically compensated; Because the consistency between the corrected simulation prediction and the physical reality is improved by more than 30%, the early warning threshold adjustment strategy is based on a higher accuracy of structural behavior prediction, enhancing the engineering credibility and field applicability of the threshold dynamic optimization process.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A stress-sensing-based disc-lock scaffolding safety early warning system, characterized in that, include: The construction phase perception module is used to acquire and determine the current construction phase information of the disc-lock scaffolding; The early warning effectiveness evaluation module is communicatively connected to stress sensors deployed at key nodes of the disc-lock scaffolding. It is used to generate effectiveness evaluation data for the current early warning threshold based on historical and real-time stress data and early warning history data collected by the stress sensors. The adaptive optimization decision module has a built-in reinforcement learning model. The reinforcement learning model takes the current construction stage information and the performance evaluation data as state inputs, and after calculation by the decision network, outputs the early warning threshold adjustment strategy for different monitoring areas. The early warning threshold dynamic execution module is connected to the adaptive optimization decision module and the early warning alarm module respectively, and is used to adjust the strategy according to the early warning threshold and update the corresponding early warning threshold in the early warning alarm module in real time. The early warning alarm module is used to compare the real-time stress data collected by the stress sensor with the updated early warning threshold. If the threshold is exceeded, a safety early warning signal is triggered.

2. The system according to claim 1, characterized in that, The construction phase sensing module further includes: The visual progress recognition unit is used to collect image data through cameras deployed at the construction site and to identify the appearance, movement and distribution of key construction components based on computer vision algorithms. The digital twin automatic update unit is connected to the visual progress recognition unit and is used to automatically drive the component status and progress plan labels in the BIM digital twin model to be updated synchronously according to the recognition results. The current construction phase information is automatically determined by the BIM digital twin model based on the updated progress status.

3. The system according to claim 2, characterized in that, The early warning effectiveness evaluation module is further integrated with: The causal inference analysis unit is used to construct a causal graph model based on historical stress data and alarm records, and to quantitatively analyze the strength of the causal relationship between specific construction activities, environmental changes and abnormal stress fluctuations. The Bayesian dynamic evaluation unit receives the output of the causal inference analysis unit and constructs the performance evaluation data into a probability distribution that changes over time. The prior probability is set based on the historical false alarm / false alarm rate, and the posterior probability is dynamically updated based on the real-time alarm results and on-site verification feedback.

4. The system according to claim 3, characterized in that, The reinforcement learning model in the adaptive optimization decision module is a multi-agent reinforcement learning architecture, wherein: The first intelligent agent makes decisions with the goal of maximizing the overall structural safety factor. The second intelligent agent makes decisions with the goal of minimizing interference with construction efficiency. Two intelligent agents negotiate strategies and conduct collaborative training through a cooperative game framework in game theory. The early warning threshold adjustment strategy is the Pareto optimal solution negotiated by both parties.

5. The system according to claim 4, characterized in that, The early warning threshold dynamic execution module further includes: The federated learning aggregation unit is used to periodically receive locally optimized threshold adjustment model parameters from cloud edge gateways of multiple similar projects while ensuring the data privacy of each construction site, and to perform secure aggregation to generate a globally optimized model. The dynamic threshold simulation and deduction unit, based on the aggregated model and the current digital twin of the disc-lock scaffolding, performs short-term stress response simulation on the threshold to be adjusted, predicts the alarm modes that may be triggered after the adjustment, and if the simulation shows that it may cause an alarm storm or cover up the real risk, it triggers the adjustment strategy rollback and manual review request.

6. The system according to claim 5, characterized in that, The early warning and alarm module further includes: The multimodal fusion alarm unit receives simultaneously a data stream from a stress sensor, a vibration / tilt data stream from an environmental sensor, and a real-time image stream from the visual progress recognition unit described in claim 2 at its input terminal. An attention-based alarm decision engine is used to perform weighted fusion analysis on the above multimodal data streams. The highest level of deterministic alarm is triggered only when evidence from multiple modalities points to the same high-risk event. If only a single modality is abnormal, a lower level of suggestive check alarm is triggered.

7. The system according to claim 6, characterized in that, The multi-agent reinforcement learning architecture also integrates: The swarm intelligence optimizer uses the ant colony algorithm to simulate multiple virtual agents exploring in parallel in the policy space. Each agent uses the initial negotiation results of the cooperative game framework as the initial pheromone distribution to perform a local fine-grained search. The dynamic environment adaptability assessment unit constructs a dynamic environment complexity index based on real-time meteorological data, equipment operating status, and worker behavior pattern data. When the index exceeds a threshold, it automatically increases the exploration weight of the swarm intelligent optimizer to discover emergency strategies to adapt to sudden complex situations.

8. The system according to claim 7, characterized in that, The cooperative game framework also includes: The social sentiment computing module uses sentiment computing algorithms to analyze the decision preference patterns of construction site managers in the trade-off between safety and efficiency, based on their historical decision-making data. The strategy explanation and visualization unit transforms the decision-making logic, trade-offs, and rationality of the final strategy of each agent during the negotiation process into an understandable security briefing using natural language generation technology, and then visualizes it using augmented reality devices to help managers understand and trust the system's decision-making.

9. The system according to claim 8, characterized in that, The federated learning aggregation unit also includes: The anomaly model detection and defense module uses adversarial machine learning technology to perform feature analysis on the local models submitted by each construction site before parameter aggregation, and to detect and isolate any malicious or low-quality model updates. The adaptive weighted aggregation optimizer dynamically adjusts the weight of each construction site in the federated aggregation based on its data quality score, model update stability index, and historical contribution. The data quality score is calculated by combining the output of the anomaly model detection and defense module, the construction site supervision score, and the historical early warning accuracy.

10. A safety early warning method for disc-lock scaffolding based on stress sensing, characterized in that, The safety early warning system for disc-lock scaffolding based on stress sensing, as described in any one of claims 1 to 9, is used to implement safety early warning for disc-lock scaffolding.

Images