Real-time stress monitoring and intelligent early warning method for hanging basket construction

Abstract

This invention discloses a real-time stress monitoring and intelligent early warning method for hanging basket construction, belonging to the field of bridge construction safety monitoring technology. Addressing the problems of lagging monitoring, isolated data, and insufficient intelligent early warning in existing technologies, this method establishes a finite element digital twin model that can dynamically simulate the mechanical state of construction based on real-time data acquisition from a multi-source sensor network. A stress anomaly identification model is constructed using deep learning algorithms, achieving intelligent diagnosis by integrating simulation and measured data. Furthermore, relying on a cloud-edge collaborative architecture, preliminary data processing and rapid alarm are performed at the edge, while high-fidelity simulation and in-depth analysis are conducted in the cloud. Based on the comprehensive evaluation results, a graded early warning information including anomaly location, risk level, and handling suggestions is generated. This method is mainly used in bridge hanging basket construction operations to achieve continuous, real-time, and intelligent safety monitoring of the stress state of its temporary load-bearing structure, effectively improving the risk early warning capability and safety management level of the construction process.

Description

Technical Field

[0001] This invention belongs to the field of bridge construction safety monitoring and intelligent early warning technology, specifically relating to a method for real-time stress monitoring and intelligent early warning in hanging basket construction. Background Technology

[0002] Current safety monitoring practices for hanging basket construction largely rely on manual inspections and periodic, discrete static stress tests. This approach suffers from monitoring lag and data discontinuity, making it difficult to capture the instantaneous stress state of the structure under dynamic changes in construction loads in real time. Because the hanging basket is a temporary spatial load-bearing structure, its stress state is complex, influenced by multiple factors such as construction procedures, material properties, and ambient temperature. Traditional monitoring methods acquire limited data points, making it difficult to construct a continuous stress history covering the entire lifecycle of key components. Consequently, it is impossible to identify and warn of potential overload, localized damage, or progressive failure risks in the early stages of construction.

[0003] Furthermore, even when online stress monitoring systems are used in some projects, their data processing and analysis methods have significant limitations. These systems typically rely solely on single-point alarm logic with fixed thresholds, lacking the ability to effectively integrate and collaboratively analyze multi-source, heterogeneous monitoring data. Simultaneously, the systems struggle to deeply integrate real-time monitoring data with simulation models reflecting the overall mechanical behavior of the structure. The lack of a digital model that evolves synchronously with the physical structure and accurately maps the boundary conditions of different construction scenarios leads to isolated and one-sided interpretations of monitoring data. This makes it impossible to distinguish between normal load fluctuations and early signs of structural anomalies, and also difficult to predict the development trend of anomalies, resulting in insufficient accuracy and foresight in early warning systems.

[0004] Furthermore, the processing and analysis of on-site monitoring data largely rely on remote central servers. This results in long data transmission chains and high demands on network stability. In the complex electromagnetic environment of construction sites and under conditions of occasional network outages, data transmission delays or even data loss can easily occur, affecting the real-time performance of monitoring. Simultaneously, the remote transmission of massive amounts of high-frequency monitoring data also brings bandwidth pressure and cost issues. How to achieve reliable data acquisition, preliminary anomaly assessment, and preservation of critical data in on-site environments with limited network conditions is a practical challenge in building a highly available real-time monitoring and early warning system.

[0005] Finally, existing technologies lack in-depth utilization of monitoring data and the ability for system self-evolution. Monitoring systems typically rely on fixed, pre-defined judgment rules, making it difficult to adapt to different projects, structural forms, or unforeseen anomalies that may occur during construction. The system cannot continuously learn from historical data and new anomaly cases, limiting its ability to cope with complex and ever-changing construction safety risks. Therefore, a monitoring and early warning method is urgently needed to address these shortcomings of existing technologies. Summary of the Invention

[0006] This invention provides a real-time stress monitoring and intelligent early warning method for hanging basket construction, which can realize real-time, continuous and intelligent monitoring and early warning of the structural stress state during the hanging basket construction process. It overcomes the problems of high lag, isolated data and insufficient intelligence of traditional monitoring methods, and significantly improves the ability to identify and respond to construction safety risks.

[0007] To achieve these objectives and other advantages of the present invention, a method for real-time stress monitoring and intelligent early warning during hanging basket construction is provided, comprising: S1. A multi-type sensor network is deployed on the key load-bearing structure of the hanging basket. The sensor network includes at least strain sensors, tilt sensors, temperature sensors and force sensors, which are used to synchronously collect structural stress, deformation posture, ambient temperature data and key point reaction force data. S2. Establish a finite element digital twin model of the hanging basket. The finite element digital twin model dynamically simulates the mechanical state of the hanging basket under the current construction load and historical load sequence based on the design drawings, material properties and real-time sensor data of the hanging basket. S3. Construct a stress anomaly identification model based on deep learning. The input of the stress anomaly identification model is the real-time simulation output of multi-source time-series sensing data and digital twin model after preprocessing and feature fusion. The output is the comprehensive stress state evaluation index and anomaly probability of key structural parts. S4. Deploy a cloud-edge collaborative computing architecture, wherein edge computing nodes are deployed at the site of the hanging basket to perform real-time filtering, compression, and preliminary anomaly judgment on sensor data, and upload the processed data to the cloud server; after receiving the data, the cloud server drives the digital twin model to perform high-fidelity simulation and calls the stress anomaly identification model for in-depth analysis. If the comprehensive stress state assessment index exceeds the preset threshold or the anomaly probability exceeds the warning value, a graded early warning information is generated; the graded early warning information is pushed to the construction management terminal in real time. The graded early warning information includes at least the abnormal location, the degree of stress exceeding the limit, the possible failure mode, and the handling suggestions.

[0008] Preferably, in step S1, the deployment of multiple sensor networks on the key load-bearing structure of the hanging basket specifically includes: The strain sensors are fixed by welding or bonding at the main truss of the hanging basket, the front / rear suspension points of the bottom basket, and the anchoring points of the traveling track to form multiple stress monitoring sections. The tilt sensor is installed at the cantilever end of the hanging basket and at the top node of the main truss; The temperature sensor is placed near the strain sensor to measure the local temperature of the structure; All sensors are protected by a waterproof and shockproof housing and are connected to the edge computing node via an industrial bus.

[0009] Preferably, in step S2, establishing a finite element digital twin model of the hanging basket specifically includes: S21. Parametric Model Construction: Based on the design drawings of the hanging basket and the mechanical property parameters of the component materials, a parametric finite element digital twin model is generated in the cloud server by using a parametric script to drive the finite element analysis kernel. S22. Construction State Driven and Dynamic Load Mapping: Establish a digital state machine that is linked to the construction progress. Based on the received process transition signals, the digital state machine drives the finite element digital twin model to automatically switch to the corresponding predefined working condition mode and update the boundary conditions. In the concrete pouring process, the finite element digital twin model dynamically calculates and applies the distributed load of wet concrete based on the position information of the concrete placing boom. At the same time, it integrates the measured reaction force data from the force sensor at the bottom basket suspension point through parallel channels as a verification and supplementary input for the distributed load theory. During the traveling operation of the hanging basket, the finite element digital twin model automatically releases the rear anchor constraint and maps the jacking force and friction load in real time based on the data from the hydraulic system sensors. S23. Model Output Verification and Confidence Management: After each simulation calculation step, the finite element digital twin model compares the theoretical stress and strain values ​​of the key parts output by the model with the measured values ​​of the strain sensors at the corresponding locations in real time, and calculates and updates a global model confidence index accordingly.

[0010] Preferably, step S23 is performed using an adaptive multi-source data fusion engine, which specifically includes: a. Input preprocessing and credibility assessment: Calculate a dynamic credibility coefficient Ci(t) for each input data source; For simulation data from the finite element digital twin model, its confidence coefficient Cm(t) is directly assigned by the global model confidence index calculated in step S23; The reliability coefficient Csj(t) of the measured data from the j-th physical sensor is calculated by taking into account the variance of the sensor’s short-term historical data, the consistency with the readings of its spatially adjacent sensors, and its own health status score. b. Dynamic weight allocation and fusion calculation: For any key part k to be evaluated on the structure, its fusion weight Wik(t) at time t is dynamically allocated according to the rule Wik(t) = Ci(t) / Σ Ci(t) based on the credibility coefficient Ci(t) of each data source, and then the final fusion stress estimate Fk(t) = Σ[Wik(t) × Xik(t)] is calculated for this part, where Xik(t) is the stress estimate provided by the i-th data source; c. Anomaly Handling and Data Reconstruction: Continuously monitor the reliability coefficient Csj(t) and jump rate of each sensor data. When a sensor s data is determined to be abnormal, the data reconstruction mechanism is automatically triggered to increase the fusion weight of the simulation data at the corresponding location. Based on the data of spatially adjacent effective sensors, the supplementary estimate of the location is reconstructed through spatial interpolation algorithm and input into the fusion calculation.

[0011] Preferably, in step S3, constructing a deep learning-based stress anomaly recognition model specifically includes: S31. Model training phase: Supervised training of deep neural network is carried out using labeled multi-source sensor data from historical projects. The labels of multi-source sensor data include at least several working conditions such as normal, local overload, anchor loosening and structural cracking. The stress anomaly identification model composed of the deep neural network is trained to output anomaly type classification and overall anomaly probability. S32. Online monitoring and unknown pattern detection stage: The real-time data stream is input into the stress anomaly identification model for inference. At the same time, the deep feature representation inside the stress anomaly identification model is extracted, and the deep feature is compared online with the pre-stored normal working condition feature library. When the mathematical distance between the deep feature and all feature clusters in the normal feature library exceeds the dynamic threshold, and the confidence of the stress anomaly identification model in classifying the currently input known anomaly type is lower than the preset threshold, it is determined to be a potential unknown anomaly mode, and the corresponding data segment is stored in the suspicious mode library. S33. Model Incremental Evolution Stage: After the suspicious pattern library has accumulated to a preset size, new pattern samples confirmed and labeled by experts are merged with historical training data to form an augmented dataset. The importance weight of each parameter in the deep neural network of the stress anomaly identification model to the historical task is calculated, and constraints proportional to this importance are applied during the optimization process for the new data to carry out incremental training of the model, so as to achieve knowledge retention and iterative updates.

[0012] Preferably, the graded early warning information is divided into at least three levels, specifically: Level 1 warning is the attention level, which corresponds to the comprehensive stress state assessment index exceeding 80% of the theoretical reference value but not reaching the design value, or the probability of abnormality being in the range of 30%-60%. The handling recommendation is to strengthen manual inspection and data monitoring. Level 2 warning is a warning level, corresponding to the comprehensive stress state assessment index reaching or exceeding the design value but not reaching the safety factor tolerance value, or the abnormal probability being in the range of 60%-90%. The handling recommendation is to suspend the current loading or moving operation and conduct a comprehensive inspection. A Level 3 warning is an alarm level, corresponding to a comprehensive stress state assessment index exceeding the allowable value of the safety factor, or an abnormal probability exceeding 90%. The recommended response is to immediately organize personnel to evacuate the danger zone and activate the emergency rescue plan.

[0013] Preferably, the cloud-edge collaborative computing architecture specifically includes a data preprocessing and fast response mechanism at the edge layer, which is implemented as follows: Edge hardware configuration and data interface: Embedded edge computing nodes are deployed at each key monitoring section of the hanging basket body. The edge computing nodes adopt low-power microprocessors and integrate multi-channel analog-to-digital converters and industrial fieldbus interfaces to directly connect to and poll the raw signals of strain sensors, tilt sensors and temperature sensors. Local rapid anomaly diagnosis rules: Each edge computing node runs a lightweight diagnostic program, which maintains a short-term historical data buffer based on a time window for each sensor channel. The lightweight diagnostic program calculates the deviation rate between the current sampled value of the sensor and the moving average of the buffer data for that channel in real time, using the formula: Deviation rate = |Current value - Moving average| / Moving standard deviation. When the deviation rate of any sensor channel continuously exceeds a preset first-level static threshold, it is determined as a potential slow anomaly. When the deviation rate instantaneously exceeds a preset second-level dynamic threshold, it is determined as a sudden and abrupt anomaly. For sudden and abrupt anomalies, the edge computing node immediately generates a lightweight alarm signal, which includes the abnormal sensor ID, anomaly type code, and a timestamp, and simultaneously triggers two independent communication links: the first link uploads data to the cloud server in real time via a low-power wide area network or a 5G slice network; the second link broadcasts data to the monitoring personnel's mobile terminal at the construction site via a local area network or Bluetooth. Data compression and asynchronous upload mechanism: In non-alarm state, the edge computing node performs lossy compression and downsampling preprocessing on the raw high-frequency sensor data collected; the rotating door trend compression algorithm is used to preserve data envelopment characteristics and reduce the sampling frequency from the original acquisition frequency to the engineering analysis frequency that satisfies the Nyquist theorem; the processed data is encapsulated into fixed-length data packets and a frame header, CRC checksum, and edge computing node ID are added; the data packets are uploaded to the cloud in batches through asynchronous transmission mode at preset fixed time periods or when the local cache reaches the capacity threshold.

[0014] Preferably, after the lightweight alarm signal is triggered, a collaborative response and data preservation process is executed, specifically as follows: Edge-side response: The edge computing node that issues a lightweight alarm signal immediately and automatically performs the following operations: a) Increases the sampling frequency of the sensor channel related to the alarm from the basic monitoring frequency to a preset high-frequency diagnostic frequency; b) Starts a raw data buffer window with a duration of T1, during which the raw high-speed sampling data of the relevant sensors is completely stored in the local non-volatile memory. Cloud-side response: After receiving a lightweight alarm signal, the cloud server performs the following operations: a) Marks the data stream of the edge computing node as high priority and allocates dedicated computing resources for processing; b) If its analysis module determines that high-fidelity data is needed for root cause analysis, it automatically sends a data retrieval command to the edge computing node. Instruction execution: After receiving the data retrieval instruction, the edge computing node uploads the cached raw high-speed sampling data to the cloud server.

[0015] Preferably, it also includes a blockchain-based monitoring data storage and audit trail mechanism, specifically: The cloud server generates hash digests with timestamps for key audit events generated during system operation in chronological order and packages them into data blocks; the key audit events include at least: original alarm signals reported by edge computing nodes, hierarchical early warning information generated in the cloud and its handling feedback status, phased records of the global confidence index of the digital twin model, important version update records of the incremental learning of the deep learning model, and execution records of data retrieval instructions; The data blocks are synchronized to a private blockchain network composed of nodes from the construction party, the supervision party, and the owner through a consensus mechanism for distributed storage. Any participating node can perform hash verification and tamper-proof traceability on the data blocks stored on the private blockchain network according to its permissions, so as to form a reliable construction safety audit log for the entire process.

[0016] Preferably, the private blockchain network is equipped with smart contracts for automatically executing compliance procedures following an early warning response, specifically including: When the cloud server generates a Level 3 alert, the smart contract is automatically triggered; The smart contract executes the following logic: First, it forcibly pushes the warning information and the corresponding panoramic data storage block to all participating nodes and requests confirmation receipts; second, if no confirmation receipt is received from any key party node within a preset time, it automatically sends a compliance alert to the preset superior regulatory platform; finally, it records the entire timeline of this warning from its generation to the responses of all parties to the blockchain.

[0017] The present invention has at least the following beneficial effects: First, this invention integrates a multi-source sensor network, a finite element digital twin model, and deep learning anomaly detection and... The cloud-edge collaborative architecture constructs a closed-loop system from data acquisition and fusion analysis to intelligent decision-making, enabling continuous, real-time, and high-fidelity monitoring of the stress state of the hanging basket structure. This overcomes the shortcomings of traditional methods, such as monitoring lag and data isolation. Through deep integration of digital twins and measured data, the monitoring data can be interpreted more accurately, distinguishing between normal fluctuations and abnormal signs. Combined with deep learning models, intelligent analysis and hierarchical early warning are performed, improving the accuracy, foresight, and automation of early warnings, thus providing strong technical support for construction safety.

[0018] Secondly, by clearly defining the core stress-bearing components such as the main truss of the hanging basket, the bottom basket lifting points, the anchoring points of the traveling track, and the cantilever... Sensors were deployed at critical deformation points such as the ends and the top of the main truss to ensure that the monitoring network could comprehensively cover the most dangerous and representative areas of the structure. The sensors were fixed by welding / bonding and fitted with protective housings to guarantee the long-term stability and reliability of the data acquisition. This deployment scheme enabled the collected data to accurately reflect the overall and local mechanical behavior of the structure, laying a solid data foundation for subsequent precise analysis and reliable early warning.

[0019] Third, by constructing a parameterized digital twin model and linking it in real time with the construction progress (digital state machine) and dynamic loads (such as concrete distribution and jacking force), the model can accurately simulate the structural response under different construction conditions. By combining the real-time comparison and confidence management of the model output with the measured data, dynamic calibration and synchronous evolution of the digital model and the physical entity are achieved.

[0020] Fourth, the adaptive multi-source data fusion engine dynamically evaluates the reliability of simulation data and data from various sensors, and dynamically allocates fusion weights accordingly, achieving complementary advantages. When a sensor malfunctions, it can automatically trigger a data reconstruction mechanism, supplementing the data with data from adjacent sensors and simulation data. This ensures the continuity of the evaluation, improves the robustness and fault tolerance of stress state estimation in key areas, and reduces the risk of misjudgment or omission due to single-point sensor failure or data anomalies, making the output of the entire monitoring system more stable and reliable.

[0021] Fifth, the deep learning-based stress anomaly identification model can not only identify known anomaly patterns (such as overload, loosening) In addition to its dynamic capabilities, the system also possesses the ability to detect unknown anomaly patterns online. Through an incremental learning mechanism, it continuously absorbs new knowledge and optimizes its models, breaking through the limitations of traditional rule-based systems and enabling the early warning system to have intelligent identification and self-evolution capabilities. The system can continuously learn from historical and real-time data, adapting to different project characteristics and new risks, thereby continuously improving its ability to detect and warn of complex and ever-changing construction safety hazards.

[0022] Sixth, a clear three-tiered early warning mechanism transforms abstract stress indicators or anomaly probabilities into concrete and actionable ones. The system provides early warning levels and response recommendations. This enables refined management of early warning information, avoiding potential false alarms or insufficient responses caused by a "one-size-fits-all" approach. Different warning levels correspond to different levels of response measures, guiding on-site personnel to take actions commensurate with the degree of risk. This ensures safety while minimizing disruption to normal construction, thus improving the scientific nature and efficiency of safety management.

[0023] Seventh, the edge layer implementation of the cloud-edge collaborative architecture is achieved by deploying embedded edge computing nodes on-site at the hanging basket. This system enables localized data collection, preprocessing, and initial anomaly diagnosis, significantly reducing reliance on cloud network stability and bandwidth, and improving system availability in harsh field environments. Local rapid diagnostic rules can detect sudden anomalies and issue local alerts immediately, ensuring timely response. Data compression and asynchronous upload mechanisms effectively reduce data transmission volume and costs, balancing data detail preservation with transmission efficiency.

[0024] Eighth, the collaborative response and data preservation process automatically increases the sampling rate and caches high-fidelity raw data at the edge when a sudden anomaly is detected. Simultaneously, the cloud can mark high-priority data and retrieve it on demand, fully preserving the raw data details from the most critical time periods before and after the anomaly. This provides invaluable high-quality data support for subsequent in-depth root cause analysis, incident review, and accountability determination. This mechanism ensures that the most valuable data is not lost or ignored due to transmission delays or routine data processing during anomaly events.

[0025] Ninth, the blockchain-based monitoring data storage and audit trail mechanism hashes key data and events (such as alarms, warnings, and model update records) and uploads them to the blockchain, forming a distributed and tamper-proof record. This fundamentally guarantees the authenticity and integrity of the monitoring data, preventing it from being tampered with or forged afterward. Any participating party (construction, supervision, owner) can conduct independent audits and traceability based on their permissions, which greatly enhances the credibility and transparency of the entire monitoring process and results, providing an authoritative and objective chain of technical evidence for resolving potential safety liability disputes.

[0026] Tenth, a compliance process for early warning response based on blockchain smart contracts, which codifies the response rules for high-level early warnings. Automation and mandatory measures ensure that critical early warning information is delivered to all relevant parties and confirmed, achieving closed-loop management of early warning response. Smart contracts automatically monitor and report response timeouts, effectively preventing response delays caused by human negligence or poor communication, strengthening the security responsibility awareness of all relevant parties, and making the entire early warning response process more standardized, efficient, and non-repudiable.

[0027] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the real-time stress monitoring and intelligent early warning method for hanging basket construction according to the present invention. Figure 2 This is a framework diagram of the real-time stress monitoring and intelligent early warning system of the present invention. Detailed Implementation

[0029] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0030] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0031] like Figure 1 As shown, this invention provides a method for real-time stress monitoring and intelligent early warning during hanging basket construction, comprising: S1. A multi-type sensor network is deployed on the key load-bearing structure of the hanging basket. The sensor network includes at least strain sensors, tilt sensors, temperature sensors and force sensors, which are used to synchronously collect structural stress, deformation posture, ambient temperature data and key point reaction force data. S2. Establish a finite element digital twin model of the hanging basket. The finite element digital twin model dynamically simulates the mechanical state of the hanging basket under the current construction load and historical load sequence based on the design drawings, material properties and real-time sensor data of the hanging basket. S3. Construct a stress anomaly identification model based on deep learning. The input of the stress anomaly identification model is the real-time simulation output of multi-source time-series sensing data and digital twin model after preprocessing and feature fusion. The output is the comprehensive stress state evaluation index and anomaly probability of key structural parts. S4. Deploy a cloud-edge collaborative computing architecture, wherein edge computing nodes are deployed at the site of the hanging basket to perform real-time filtering, compression, and preliminary anomaly judgment on sensor data, and upload the processed data to the cloud server; after receiving the data, the cloud server drives the digital twin model to perform high-fidelity simulation and calls the stress anomaly identification model for in-depth analysis. If the comprehensive stress state assessment index exceeds the preset threshold or the anomaly probability exceeds the warning value, a graded early warning information is generated; the graded early warning information is pushed to the construction management terminal in real time. The graded early warning information includes at least the abnormal location, the degree of stress exceeding the limit, the possible failure mode, and the handling suggestions.

[0032] In the above embodiments, firstly, a multi-sensor network is deployed on the key load-bearing structure of the hanging basket. The core of this sensor network lies in integrating various types of sensors, such as strain sensors capable of detecting minute material deformations and calculating stress, tilt sensors for monitoring overall or local tilt changes in the structure, and temperature sensors for collecting local ambient temperature data. These sensors are not installed arbitrarily, but rather selectively fixed to key locations with concentrated internal forces or sensitive deformation, such as the main truss, the front and rear suspension points of the bottom basket, and the anchoring points of the traveling track, based on structural mechanics analysis, forming multiple stress monitoring sections covering the main force transmission paths. All sensors are typically reliably fixed by welding or high-strength adhesives and fitted with protective housings to resist environmental impacts such as water splashes and impacts during construction. They are connected to nearby edge computing nodes via industrial fieldbuses (such as CAN bus or RS485) to achieve synchronous and high-speed data acquisition. The fundamental principle of this arrangement is that by coordinating multiple types of sensors in space, the physical state of the structure can be captured simultaneously from three dimensions: stress, deformation, and temperature. This provides a comprehensive and integrated raw data stream for subsequent analysis, overcoming the shortcomings of traditional single-point stress test data, which are one-sided and discontinuous.

[0033] Secondly, a finite element digital twin model is established to evolve synchronously with the physical formwork and to simulate its mechanical behavior with high fidelity. The construction of this finite element digital twin model begins with static information such as design drawings and material properties (e.g., elastic modulus, Poisson's ratio). A parametric script drives the finite element analysis kernel to automatically generate the basic model in the cloud. Its core "dynamic" nature is reflected in its real-time linkage with the construction process: the finite element digital twin model has a built-in digital state machine that can automatically switch to preset working condition modes (e.g., "pouring condition," "walking condition") based on received process transition signals (e.g., "start pouring," "start walking"), and update the corresponding boundary constraints. For example, during concrete pouring, the finite element digital twin model can dynamically calculate and apply non-uniform wet concrete distributed loads based on the concrete placing boom's position information; simultaneously, it will also receive measured reaction force data from the force sensors at the bottom of the formwork suspension point, using these measured values ​​as verification and supplementary inputs to the theoretical distributed load, thus more realistically reflecting the actual load. During the hanging basket travel operation, the finite element digital twin model automatically releases the rear anchorage constraint and maps loads such as the jacking force and track friction force in real time based on sensor data from the hydraulic propulsion system. After each simulation step is completed, the finite element digital twin model compares the theoretical stress and strain values ​​of key parts output by the model with the measured values ​​of strain sensors at the corresponding locations in real time. Based on this difference, a global model confidence index is calculated and continuously updated. The principle of this process is to continuously verify the "physical measurement - virtual simulation" closed loop, so that the digital model continuously approximates the real state of the physical entity, forming a calculable and predictable "virtual copy".

[0034] Then, a deep learning-based stress anomaly identification model is constructed for intelligent diagnosis of the fused data. During the training phase, this model utilizes supervised learning with a large amount of multi-source time-series sensor data from historical projects, labeled with clear tags (such as "normal," "local overload," "anchor loosening," and "microcrack propagation"), to train a deep neural network (e.g., a Long Short-Term Memory network (LSTM) or a Temporal Convolutional Network (TCN)) to identify features of different anomaly patterns. In actual online monitoring, the input to the stress anomaly identification model is a multi-source real-time data stream that has undergone preprocessing and feature engineering fusion. This data stream includes not only readings from various sensors but also real-time simulation outputs from the aforementioned digital twin model, such as theoretical stress distribution. After inference, the stress anomaly identification model not only outputs a classification judgment for known anomaly types but also provides a probability value representing the overall anomaly likelihood, for example, a value between 0 and 1. Furthermore, the stress anomaly identification model also possesses the ability to detect "unknown" anomalies: it extracts deep features from the data and compares them with a large pre-stored feature library of "normal operating conditions"; if the current feature differs significantly from all known normal patterns, and the stress anomaly identification model has low confidence in classifying known anomalies, it is marked as a "potential unknown anomaly" and stored in a suspicious pattern library for subsequent expert analysis. The core principle of this part is to utilize the powerful nonlinear feature extraction and pattern recognition capabilities of deep neural networks to automatically learn and identify subtle abnormal signs from complex, high-dimensional time-series data, with a level of intelligence far exceeding simple alarm rules based on fixed thresholds.

[0035] Finally, a cloud-edge collaborative computing and decision-making architecture is deployed to enable the engineering operation of the entire method. At the hanging basket construction site, embedded edge computing nodes are deployed at each key monitoring section. These nodes are directly connected to sensors and are responsible for first-hand processing of the raw high-frequency signals, including real-time filtering to remove noise, data compression to save bandwidth (e.g., using a rotating door trend compression algorithm to significantly reduce data volume while preserving data change trends), and preliminary anomaly judgment based on simple rules (such as checking whether data changes instantaneously or continuously deviates from the historical moving average). The processed data and preliminary alarm signals are uploaded to the cloud server via a wireless network (such as 4G / 5G or a private network). The cloud server has powerful computing capabilities and, upon receiving the data, mainly performs two core tasks: first, driving the aforementioned high-fidelity finite element digital twin model for real-time simulation; and second, calling the trained deep learning stress anomaly recognition model for in-depth analysis and comprehensive judgment. The cloud analysis engine comprehensively considers the measured data, simulation results, and the anomaly probability output by the intelligent model to generate a comprehensive stress state assessment index for key structural components. This stress state assessment index can be compared with preset multi-level thresholds. For example, a warning level can be triggered when the stress state assessment index exceeds 80% of the design allowable value; a warning level can be triggered when it exceeds the design value but does not reach the safety allowable value; and the highest level alarm can be triggered when it exceeds the safety allowable value (the safety factor is usually between 1.3 and 2.0, with the specific value determined according to engineering specifications and design documents). The warning information is not just a simple "exceeding the limit" alarm, but includes the specific location of the anomaly, the severity of the stress exceeding the limit, the inferred possible failure mode (such as "risk of instability under compression of the lower chord of the main truss"), and targeted handling suggestions (such as "suspend loading and check anchor bolts"). These tiered warning messages are pushed in real time to various terminals such as handheld terminals of construction management personnel and large screens in the monitoring center to guide rapid on-site response.

[0036] Compared to existing technologies that rely on manual inspections and discrete static tests, this invention achieves continuous, high-density, and automated monitoring of the stress state of the hanging basket structure, significantly overcoming the problem of monitoring lag. Compared to systems that only use online monitoring but lack in-depth analysis, this invention introduces a digital twin model, interpreting isolated monitoring point data within the mechanical context of the entire structural system. Furthermore, it utilizes a deep learning model to achieve intelligent identification and early warning of complex anomaly patterns, qualitatively improving the accuracy, foresight, and intelligence of the warnings. In addition, the cloud-edge collaborative architecture effectively resolves the contradiction between poor network conditions at construction sites and the need for massive data transmission and real-time processing. Preprocessing and rapid judgment at the edge significantly enhance the system's robustness and real-time response capabilities in complex environments. Simultaneously, resource-intensive high-fidelity simulation and in-depth analysis in the cloud ensure the scientific rigor and reliability of the warning decisions. Overall, this method constructs a complete closed loop from data perception, fusion modeling, intelligent diagnosis to decision feedback, providing a more proactive, precise, and efficient intelligent safeguard for the safety of hanging basket construction.

[0037] In one specific embodiment, step S1, which involves deploying a multi-sensor network on the key load-bearing structure of the hanging basket, specifically includes: The strain sensors are fixed by welding or bonding at the main truss of the hanging basket, the front / rear suspension points of the bottom basket, and the anchoring points of the traveling track to form multiple stress monitoring sections. The tilt sensor is installed at the cantilever end of the hanging basket and at the top node of the main truss; The temperature sensor is placed near the strain sensor to measure the local temperature of the structure; All sensors are protected by a waterproof and shockproof housing and are connected to the edge computing node via an industrial bus.

[0038] In the above implementation, the specific deployment of multiple sensor networks on the key load-bearing structure of the hanging basket is first clarified to ensure that the monitoring data can accurately reflect the core mechanical behavior of the structure. Specifically, strain sensors are selectively arranged at key locations where internal forces are concentrated or transmitted, such as the main truss of the hanging basket, the front / rear suspension points of the bottom basket, and the anchoring points of the traveling track. These strain sensors typically employ resistance strain gauges or fiber optic grating sensors, which are fixed by welding or high-strength structural adhesive to form multiple stress monitoring sections covering the main force transmission paths, thereby achieving distributed and representative measurement of structural stress.

[0039] Secondly, to monitor the overall attitude and deformation of the formwork during construction, tilt sensors, such as MEMS tiltmeters or electrolyte-based tilt sensors, are installed at the cantilever ends and the top nodes of the main truss. The cantilever ends are where deformation is most significant, while the top nodes of the main truss reflect the overall torsional or tilting trend of the structure. These tilt sensors can output the tilt changes of the structure in real time, and their data is crucial for judging structural stability and verifying calculation models. Simultaneously, considering that temperature changes cause thermal expansion and contraction of materials, thus affecting strain measurement results, temperature sensors, such as PT100 platinum resistance thermometers or digital temperature sensors, are placed near each strain sensor to measure and compensate for local temperature effects in the structure, improving the accuracy of strain data calculation.

[0040] Finally, to ensure the long-term reliable operation of the sensors in harsh construction environments, protection and integration requirements were imposed on all sensors. Each sensor is encapsulated in an industrial-grade housing with waterproof and shockproof features to protect against rain, concrete slurry splashes, and potential physical impacts. All sensors are connected into a unified network via standard industrial buses (e.g., CAN bus, RS485 bus, or Ethernet-based industrial protocols), ultimately converging at edge computing nodes deployed in the field. This centralized, bus-based connection method simplifies wiring, improves system reliability and maintainability, and ensures that raw data can be transmitted stably and at high speed to the data processing unit.

[0041] By clearly defining the location, type, and installation method of sensors, the systematic nature, representativeness, and accuracy of monitoring data acquisition are ensured. This avoids arbitrary sensor deployment, enabling the collected data to accurately and comprehensively reflect the stress state and deformation behavior of key components of the hanging basket. This provides a high-quality, highly reliable data foundation for subsequent digital twin model calibration and intelligent analysis, fundamentally improving the reliability of the entire monitoring and early warning system.

[0042] In one specific implementation, step S2, establishing a finite element digital twin model of the hanging basket, specifically includes: S21. Parametric Model Construction: Based on the design drawings of the hanging basket and the mechanical property parameters of the component materials, a parametric finite element digital twin model is generated in the cloud server by using a parametric script to drive the finite element analysis kernel. S22. Construction State Driven and Dynamic Load Mapping: Establish a digital state machine that is linked to the construction progress. Based on the received process transition signals, the digital state machine drives the finite element digital twin model to automatically switch to the corresponding predefined working condition mode and update the boundary conditions. In the concrete pouring process, the finite element digital twin model dynamically calculates and applies the distributed load of wet concrete based on the position information of the concrete placing boom. At the same time, it integrates the measured reaction force data from the force sensor at the bottom basket suspension point through parallel channels as a verification and supplementary input for the distributed load theory. During the traveling operation of the hanging basket, the finite element digital twin model automatically releases the rear anchor constraint and maps the jacking force and friction load in real time based on the data from the hydraulic system sensors. S23. Model Output Verification and Confidence Management: After each simulation calculation step, the finite element digital twin model compares the theoretical stress and strain values ​​of the key parts output by the model with the measured values ​​of the strain sensors at the corresponding locations in real time, and calculates and updates a global model confidence index accordingly.

[0043] The above implementation details the process of establishing a finite element digital twin model. First, the construction of the finite element digital twin model begins with the construction of a parametric model. This step, based on the precise design drawings of the hanging basket and the material mechanical properties of all components, such as the elastic modulus, yield strength, and Poisson's ratio of steel, uses pre-written parametric scripts (e.g., developed using Python and finite element software APIs) on a cloud server to drive the finite element analysis kernel, such as the solver of ABAQUS or ANSYS, to automatically generate a basic digital twin model. The parametric nature of this finite element digital twin model means that its geometric dimensions, material properties, etc., can be quickly adjusted through scripts, facilitating adaptation to different projects or design changes and providing convenience for the reuse and rapid deployment of the finite element digital twin model.

[0044] Secondly, the dynamism of the finite element digital twin model is achieved through construction state-driven and dynamic load mapping. The system establishes a "digital state machine" that is linked to the construction progress in the physical world. When it receives a process transition signal from the construction management system, such as "start pouring the Nth segment" or "the formwork begins to move forward," the digital state machine drives the digital twin model to automatically switch to the corresponding predefined "working condition mode" and update the model's boundary conditions. Under the concrete pouring condition, the system acquires the three-dimensional position coordinates of the concrete placing boom in real time through a positioning module (such as GNSS or UWB) installed on the boom and transmits this coordinate information to the cloud server. The finite element digital twin model dynamically calculates and applies a non-uniform wet concrete distributed load based on this real-time boom position information. At the same time, measured reaction force data from the force sensors at the bottom basket suspension points are integrated in parallel channels as a verification and supplementary input for the above-mentioned position-based distributed load theory, thereby more realistically reflecting the actual load. During the traveling operation of the hanging basket, the finite element digital twin model automatically releases the consolidation constraints of the rear anchor point and maps the loads such as the jacking force and track friction force in real time based on the pressure sensor data installed on the hydraulic jacking device, thereby simulating the complex force state during the traveling process.

[0045] Finally, to ensure consistency between the digital model and the physical entity, model output verification and confidence management were introduced. After each simulation calculation step of the digital twin model, the system compares the theoretical stress and strain values ​​of key components (such as the mid-span of the main truss and the base of the hangers) with the real-time measured values ​​from strain sensors installed at the same locations. Based on the differences, for example using root mean square error or correlation coefficient, the system dynamically calculates and updates a global model confidence index. This global model confidence index is a value between 0 and 1. For example, when the simulated values ​​and measured values ​​are highly consistent, the confidence index can be close to 0.9 or higher; when the difference is large, the confidence index decreases. This global model confidence index intuitively reflects the accuracy of the current digital model in simulating the physical world, providing an important basis for subsequent data fusion and decision-making.

[0046] The digital twin model constructed in this implementation is not a static, offline computational model, but a "living" virtual mirror that is linked to the construction progress in real time and dynamically verified with measured data. By accurately mapping construction conditions and dynamic loads, and combining measured data for closed-loop correction, it achieves high-fidelity, real-time synchronous simulation of the mechanical state of the physical hanging basket. This greatly compensates for the shortcomings of traditional monitoring, such as isolated data interpretation and the inability to predict changes in the overall structural behavior. It enables early warning decisions to be based on a deeper and more comprehensive understanding of structural mechanics, significantly improving the scientific rigor and predictability of early warnings.

[0047] In one specific implementation, step S23 is executed through an adaptive multi-source data fusion engine, which specifically includes: a. Input preprocessing and credibility assessment: Calculate a dynamic credibility coefficient Ci(t) for each input data source, where Ci(t) is the dynamic credibility coefficient of the i-th data source at time t; For simulation data from the finite element digital twin model, its confidence coefficient Cm(t) is directly assigned by the global model confidence index calculated in step S23; where Cm(t) is the confidence coefficient of the digital twin model at time t; For measured data from the j-th physical sensor, the reliability coefficient Csj(t) is calculated by combining the variance of the sensor's short-term historical data, the consistency with the readings of its spatially adjacent sensors, and its own health status score; where Csj(t) is the reliability coefficient of the j-th physical sensor at time t. b. Dynamic Weight Allocation and Fusion Calculation: For any key component k to be evaluated on the structure, its fusion weight Wik(t) at time t is dynamically allocated based on the credibility coefficient Ci(t) of each data source according to the rule Wik(t) = Ci(t) / Σ Ci(t), and then the final fusion stress estimate of the component Fk(t) = Σ[Wik(t) × Xik(t)] is calculated; where Wik(t) is the fusion weight of the i-th data source for the k-th component; Σ Ci(t) is the sum of the credibility coefficients of all data sources, used to normalize the weights; Xik(t) is the stress estimate of the k-th component provided by the i-th data source; Fk(t) is the final fusion stress estimate of the k-th key component at time t; c. Anomaly Handling and Data Reconstruction: Continuously monitor the reliability coefficient Csj(t) and jump rate of each sensor data. When a sensor s data is determined to be abnormal, the data reconstruction mechanism is automatically triggered to increase the fusion weight of the simulation data at the corresponding location. Based on the data of spatially adjacent effective sensors, the supplementary estimate of the location is reconstructed through spatial interpolation algorithm and input into the fusion calculation.

[0048] In the above implementation, the core objective of the adaptive multi-source data fusion engine is to comprehensively utilize simulation data from the digital twin model and measured data from multiple physical sensors to obtain a more reliable and robust estimate of the stress state of key structural components. Its workflow begins with input preprocessing and confidence assessment. The adaptive multi-source data fusion engine dynamically calculates a confidence coefficient for each input data source, including the simulation data stream from the digital twin model and the measured data stream from each physical sensor. For simulation data, the confidence coefficient is directly assigned by the calculated global model confidence index; the more accurate the finite element digital twin model, the higher its data confidence. For physical sensor data, the calculation of its confidence coefficient comprehensively considers three factors: the fluctuation of the sensor's short-term historical data (smaller variance indicates greater stability and higher confidence), the consistency of readings from other sensors adjacent to its installation location (higher consistency indicates higher confidence), and the sensor's own health status score (such as power supply voltage, communication quality, etc.).

[0049] Next, the adaptive multi-source data fusion engine performs dynamic weight allocation and fusion calculation. For any critical structural component to be evaluated, such as the midpoint of a main truss member, the engine collects all data sources that can provide stress estimates for that component. At each calculation step, the engine dynamically allocates fusion weights proportionally based on the dynamic reliability coefficients of each data source calculated in the previous step. A simple and effective rule is that the higher the reliability coefficient of a data source, the greater its weight in the stress fusion estimate for that component. Based on the allocated weights, the engine calculates the final "fusion stress estimate" for that component through a weighted summation. This process ensures that high-reliability data dominates the decision-making process.

[0050] Finally, the adaptive multi-source data fusion engine possesses intelligent fault-tolerant capabilities for anomaly handling and data reconstruction. The engine continuously monitors the reliability coefficient and data jump rate of each physical sensor's data, i.e., instantaneous abnormal changes. Once a sensor is determined to have malfunctioned or output abnormal data—for example, its reliability coefficient consistently falls below a threshold, such as 0.3, or its reading experiences an impossible drastic jump—the engine automatically triggers a data reconstruction mechanism. At this time, the system proactively reduces the fusion weight of the abnormal sensor's data while simultaneously increasing the fusion weight of the simulation data from the digital twin model at the corresponding location. Furthermore, based on readings from other spatially adjacent and still valid sensors, the system reconstructs a supplementary stress estimate for the location of the abnormal sensor using spatial interpolation algorithms (such as inverse distance weighting), and inputs this estimate as a new data source for fusion calculation. This ensures that even if individual sensors fail, the stress assessment of that location will not be interrupted or severely distorted.

[0051] Through an adaptive multi-source data fusion engine, intelligent complementarity and dynamic optimization of simulation and measured data are achieved. It no longer simply treats all data equally or takes an average, but dynamically adjusts its influence based on the real-time quality of the data source. This significantly improves the accuracy and robustness of stress state assessment in key components. In particular, its anomaly handling and data reconstruction functions endow the system with strong fault tolerance, effectively suppressing the risk of misjudgment caused by single-point sensor failures, transient interference, or data anomalies. This ensures the continuity and stability of the monitoring system output, making the foundation for early warning decisions more solid and reliable.

[0052] In one specific implementation, step S3 involves constructing a deep learning-based stress anomaly recognition model, specifically including: S31. Model training phase: Supervised training of deep neural network is carried out using labeled multi-source sensor data from historical projects. The labels of multi-source sensor data include at least several working conditions such as normal, local overload, anchor loosening and structural cracking. The stress anomaly identification model composed of the deep neural network is trained to output anomaly type classification and overall anomaly probability. S32. Online monitoring and unknown pattern detection stage: The real-time data stream is input into the stress anomaly identification model for inference. At the same time, the deep feature representation inside the stress anomaly identification model is extracted, and the deep feature is compared online with the pre-stored normal working condition feature library. When the mathematical distance between the deep feature and all feature clusters in the normal feature library exceeds the dynamic threshold, and the confidence of the stress anomaly identification model in classifying the currently input known anomaly type is lower than the preset threshold, it is determined to be a potential unknown anomaly mode, and the corresponding data segment is stored in the suspicious mode library. S33. Model Incremental Evolution Stage: After the suspicious pattern library has accumulated to a preset size, new pattern samples confirmed and labeled by experts are merged with historical training data to form an augmented dataset. The importance weight of each parameter in the deep neural network of the stress anomaly identification model to the historical task is calculated, and constraints proportional to this importance are applied during the optimization process for the new data to carry out incremental training of the model, so as to achieve knowledge retention and iterative updates.

[0053] In the above implementation, the stress anomaly identification model is the core brain of the entire system for achieving intelligent early warning. The entire process of building a deep learning-based stress anomaly identification model is divided into three stages. The first stage is the model training stage. In this first stage, it is necessary to collect a large amount of multi-source sensor data from historical hanging basket construction projects, and this data needs to be clearly labeled by domain experts according to the actual situation. The label types should at least cover a variety of typical working conditions such as "normal", "local overload", "anchor loosening" and "structural cracking". Using these labeled time-series data (including change curves of multiple dimensions such as strain, tilt angle, and temperature), a deep neural network (e.g., an LSTM network or a one-dimensional convolutional neural network that is good at processing time-series data) is trained under supervision. The training goal is to enable the model to learn to identify the corresponding anomaly type from complex data patterns and to output a probability value representing the overall probability of an anomaly.

[0054] The second phase is the online monitoring and unknown pattern detection phase. The trained stress anomaly identification model is deployed to the cloud to perform online inference on the real-time data stream, determining which known anomaly type the current state belongs to and providing an anomaly probability. It is also designed to detect "unknown" anomaly patterns. Specifically, while the stress anomaly identification model is performing inference, it extracts the activation values ​​of a hidden layer as a "deep feature representation" of the data. The system maintains a normal operating condition feature library pre-calculated on a large amount of normal operating condition data. During online monitoring, the system compares the deep features of the real-time data with all feature clusters in the feature library, calculating their mathematical distances, such as Euclidean or cosine distances. The system sets a dynamic threshold, for example, the 95th percentile of the distance can be used as a reference. If the distance between the current deep feature and all normal features exceeds this dynamic threshold, and the stress anomaly identification model has a low confidence level in classifying the current input as belonging to any known anomaly type (e.g., the confidence level for the highest category is below 50%), then the system will determine that the current state may be a potential unknown anomaly pattern. At this point, the data segment that triggered the judgment will be automatically saved to the suspicious pattern library for subsequent expert analysis.

[0055] The third stage is the incremental evolution stage of the model. As construction progresses, the suspicious pattern database will continuously accumulate data. When it reaches a certain scale, such as dozens of different suspicious segments, experts can be organized to review and manually label these segments to confirm whether they are new abnormal patterns and assign them new labels. These confirmed and labeled new pattern samples are merged with the original historical training data to form an augmented dataset. To protect the old knowledge already learned by the model and prevent "catastrophic forgetting," a strategy called elastic weight consolidation is adopted when incrementally training the model using the new dataset. This elastic weight consolidation strategy calculates the importance weight of each parameter in the original neural network for correctly classifying historical tasks, and during the optimization process for new data, imposes constraints proportional to the importance of important old parameters, limiting their significant changes. In this way, the stress anomaly recognition model can both learn new knowledge and retain old knowledge well, achieving continuous accumulation and iterative updates of knowledge.

[0056] The constructed deep learning model not only achieves accurate identification of known risks, but more importantly, endows the system with the "sense" to detect unknown risks, as well as the intelligence to continuously learn and evolve. This enables the monitoring and early warning system to break through the limitations of preset rules and proactively adapt to the specificities of different engineering projects and new problems and risks that may arise during construction. The system has moved from "experience solidification" to "experience growth," achieving a qualitative leap in its intelligence level and adaptive capabilities, enabling it to cope with more complex and ever-changing construction safety challenges.

[0057] In one specific implementation, the graded early warning information is divided into at least three levels, specifically: Level 1 warning is the attention level, which corresponds to the comprehensive stress state assessment index exceeding 80% of the theoretical reference value but not reaching the design value, or the probability of abnormality being in the range of 30%-60%. The handling recommendation is to strengthen manual inspection and data monitoring. Level 2 warning is a warning level, corresponding to the comprehensive stress state assessment index reaching or exceeding the design value but not reaching the safety factor tolerance value, or the abnormal probability being in the range of 60%-90%. The handling recommendation is to suspend the current loading or moving operation and conduct a comprehensive inspection. A Level 3 warning is an alarm level, corresponding to a comprehensive stress state assessment index exceeding the allowable value of the safety factor, or an abnormal probability exceeding 90%. The recommended response is to immediately organize personnel to evacuate the danger zone and activate the emergency rescue plan.

[0058] In the above implementation, the early warning information is defined in a hierarchical manner, establishing a clear three-level early warning system, transforming abstract monitoring indicators into specific and actionable guidelines. The first level is the attention-level early warning. The triggering conditions for this level of early warning are relatively lenient. For example, when the "comprehensive stress state assessment index" calculated by the system exceeds 80% of the theoretical reference value for that part, but has not yet reached the design value, or when the "anomaly probability" output by the deep learning model is in a low-risk range, such as between 30% and 60%. At this time, although there is no immediate danger to structural safety, a noteworthy trend has emerged. Therefore, the corresponding "handling recommendations" focus on observation and enhanced monitoring, such as "increasing the frequency of manual inspections" and "focusing on monitoring relevant data changes at the monitoring center," with the aim of reminding relevant personnel to be vigilant.

[0059] Level 2 is a warning level. This level indicates a significantly increased risk, approaching or reaching the design safety boundary. The trigger condition is that the comprehensive stress state assessment index reaches or exceeds the "design value" (i.e., the theoretical bearing capacity of the structure), but has not yet reached the "safety factor tolerance value" after considering the safety factor, for example, the design value multiplied by a reduction factor less than 1, usually determined based on specifications, possibly between 0.6 and 0.8. Alternatively, the probability of anomaly is in a high range, such as between 60% and 90%. At this point, the structure may be on the verge of being unsafe or about to become unsafe. Therefore, the recommended action is to escalate to direct intervention, the core of which is to suspend current loading or movement operations to prevent further risk aggravation, and immediately initiate a comprehensive inspection to find the specific cause of the stress anomaly, such as checking for loose bolts, cracked welds, and excessive loads.

[0060] Level 3 is the highest level of alarm warning. This level indicates that the structure is in a highly dangerous state with an extremely high risk of failure. The trigger condition is that the comprehensive stress state assessment index exceeds the allowable value of the safety factor, or the probability of anomaly exceeds 90%. The allowable value of the safety factor is the final safety baseline after considering factors such as material uncertainties and construction errors. Once exceeded, the possibility of structural instability or failure increases dramatically. At this point, any continuation of construction or hesitation could lead to an accident. Therefore, the recommended response is a mandatory safety emergency measure: "Immediately organize the evacuation of personnel from the danger zone," prioritizing personnel safety, while simultaneously "activating the emergency rescue plan," mobilizing resources for emergency reinforcement or other rescue measures.

[0061] It should be noted that the threshold benchmarks for each level of early warning are defined as follows: Theoretical reference value: refers to the theoretical stress or strain value of key parts of the structure under no-load or design standard load, calculated by the finite element digital twin model relative to the current construction conditions. This value serves as the benchmark for determining whether the monitoring data deviates from the "normal" theoretical state. Design value: refers to the allowable stress of the material or the allowable deformation of the structure specified in the design document of the hanging basket structure. Allowable safety factor value: refers to the final control value after considering material uncertainties, construction errors, and other safety factors based on the design value, usually the design value divided by a safety factor greater than 1 (determined according to engineering specifications).

[0062] By clearly defining a three-tiered early warning system, refined and differentiated management of early warning information is achieved. It maps continuous and complex technical indicators to discrete and clear action instructions, providing clear and direct decision support for management and operational personnel at different levels.

[0063] In one specific implementation, the cloud-edge collaborative computing architecture specifically includes a data preprocessing and fast response mechanism at the edge layer, which is implemented as follows: Edge hardware configuration and data interface: Embedded edge computing nodes are deployed at each key monitoring section of the hanging basket body. The edge computing nodes adopt low-power microprocessors and integrate multi-channel analog-to-digital converters and industrial fieldbus interfaces to directly connect to and poll the raw signals of strain sensors, tilt sensors and temperature sensors. Local rapid anomaly diagnosis rules: Each edge computing node runs a lightweight diagnostic program, which maintains a short-term historical data buffer based on a time window for each sensor channel. The lightweight diagnostic program calculates the deviation rate between the current sampled value of the sensor and the moving average of the data in the buffer for that channel in real time, using the formula: Deviation rate = |Current value - Moving average| / Moving standard deviation. When the deviation rate of any sensor channel continuously exceeds a preset first-level static threshold, it is determined as a potential slow anomaly. When the deviation rate instantaneously exceeds a preset second-level dynamic threshold, it is determined as a sudden and abrupt anomaly, where the value of the second-level dynamic threshold is greater than the first-level static threshold. For sudden and abrupt anomalies, the edge computing node immediately generates a lightweight alarm signal, which includes the abnormal sensor ID, anomaly type code, and a timestamp, and simultaneously triggers two independent communication links: the first link uploads data to the cloud server in real time via a low-power wide area network or a 5G slice network; the second link broadcasts data to the monitoring personnel's mobile terminal at the construction site via a local area network or Bluetooth. Data compression and asynchronous upload mechanism: In non-alarm state, the edge computing node performs lossy compression and downsampling preprocessing on the raw high-frequency sensor data collected; the rotating door trend compression algorithm is used to preserve data envelopment characteristics and reduce the sampling frequency from the original acquisition frequency to the engineering analysis frequency that satisfies the Nyquist theorem; the processed data is encapsulated into fixed-length data packets and a frame header, CRC checksum, and edge computing node ID are added; the data packets are uploaded to the cloud in batches through asynchronous transmission mode at preset fixed time periods or when the local cache reaches the capacity threshold.

[0064] The above implementation details the hardware configuration, data processing, and rapid response mechanism of the "edge layer" in the cloud-edge collaborative architecture. Firstly, at the hardware level, this method requires the deployment of embedded edge computing nodes at key monitoring sections of the hanging basket body. These embedded edge computing nodes are not ordinary microcontrollers, but dedicated low-power microprocessor modules integrating multi-channel analog-to-digital converters (ADCs) and industrial fieldbus interfaces (such as CAN controllers and RS485 transceivers), for example, based on the ARM Cortex-M series core. They are directly mounted on the hanging basket structure and connected to all strain, tilt, and temperature sensors at that section via cables, enabling direct, polling-style acquisition of the raw analog or digital signals output by these sensors, ensuring first-hand data acquisition and low latency.

[0065] Secondly, each edge computing node internally runs a lightweight diagnostic program. This program maintains a short-term historical data buffer for each connected sensor channel, based on a time window (e.g., the most recent 10 or 30 seconds). The core logic of the lightweight diagnostic program is to calculate in real-time the deviation rate between the current sampled value of each sensor and the moving average of its buffer data. The deviation rate is calculated by dividing the absolute difference between the current value and the moving average by the standard deviation of the data within the moving window. This deviation rate effectively measures the degree of deviation of the current reading from the recent historical normal fluctuation range. The lightweight diagnostic program presets two levels of judgment thresholds. The first level is a lower static threshold used to detect slowly developing abnormal trends. Its judgment logic is based on persistent deviations. For example, when the deviation rate calculated by a sensor channel (i.e., the standardized deviation of the current sampled value from the moving average) continuously exceeds a preset lower threshold, for example, a deviation rate > 3, and continues to reach a preset number of periods, such as 5 consecutive sampling periods, it is judged as a potential slow anomaly. The second level is a higher dynamic threshold specifically used to detect sudden and drastic changes. Its judgment logic is based on instantaneous sharp deviations. For example, when the deviation rate instantaneously exceeds a preset high threshold, such as a deviation rate > 10, it is immediately judged as a sudden and drastic anomaly. Here, the threshold (e.g., 10) means that the current sampled value deviates from the moving average by more than 10 times the recent normal fluctuation range of the sensor data (i.e., its moving standard deviation). In the event of a sudden and drastic anomaly, the edge computing node immediately generates a lightweight alarm signal containing the sensor ID, anomaly type code, and timestamp. To ensure the absolute reliability of the alarm, the signal is sent simultaneously through two independent communication links: the first link uploads the signal to the cloud server in real time via a low-power wide area network (such as NB-IoT) or a highly reliable 5G slice network; the second link broadcasts the signal directly to the mobile terminal of the monitoring personnel at the construction site via local Wi-Fi, LAN, or Bluetooth. This dual-link redundancy mechanism ensures that even in the extreme case of a temporary interruption of the cloud network, on-site personnel can receive critical alarm information immediately.

[0066] Finally, to save network bandwidth and cloud storage costs, edge computing nodes perform lossy compression and downsampling preprocessing on the collected high-frequency raw sensor data, such as the raw sampling rate of 100Hz. Specifically, a rotating door trend compression algorithm can be used, which effectively preserves the overall trend and key turning points of the data while significantly reducing the number of data points. Subsequently, the data sampling frequency is reduced from the original acquisition frequency to a frequency that meets the basic requirements of engineering analysis, such as 20Hz or 10Hz, satisfying the Nyquist theorem, i.e., more than twice the highest frequency component of interest. The processed data is encapsulated into fixed-length data packets with a frame header, CRC checksum, and the local node ID. These data packets are not uploaded in real time, but are uploaded to the cloud asynchronously and in batches at preset fixed intervals (e.g., every 30 seconds) or when the local cache reaches a certain capacity, thereby significantly reducing the average network load.

[0067] By performing intelligent preprocessing and preliminary judgment at the data source, the core challenges of poor network conditions and unstable data transmission at construction sites are effectively solved. This endows the system with strong local autonomy and rapid response capabilities. In particular, the local real-time diagnosis and dual-link alarm for sudden anomalies minimize warning delays, buying valuable time for personnel to evacuate in emergencies. Simultaneously, its data compression and asynchronous upload mechanism reduces network bandwidth pressure and cloud processing burden while ensuring no loss of critical information, achieving a good balance between real-time monitoring, reliability, and economy. This makes the intelligent monitoring system highly practical and usable in complex engineering sites.

[0068] In one specific implementation, after the lightweight alarm signal is triggered, a collaborative response and data preservation process is executed, specifically as follows: Edge-side response: The edge computing node that issues a lightweight alarm signal immediately and automatically performs the following operations: a) Increases the sampling frequency of the sensor channel related to the alarm from the basic monitoring frequency to a preset high-frequency diagnostic frequency; b) Starts a raw data buffer window with a duration of T1, during which the raw high-speed sampling data of the relevant sensors is completely stored in the local non-volatile memory. Cloud-side response: After receiving a lightweight alarm signal, the cloud server performs the following operations: a) Marks the data stream of the edge computing node as high priority and allocates dedicated computing resources for processing; b) If its analysis module determines that high-fidelity data is needed for root cause analysis, it automatically sends a data retrieval command to the edge computing node. Instruction execution: After receiving the data retrieval instruction, the edge computing node uploads the cached raw high-speed sampling data to the cloud server.

[0069] In the above implementation, after a lightweight alarm signal is triggered, the system collaboratively executes a response and data preservation process on both the edge and cloud sides, aiming to capture and save the most detailed data before and after the abnormal event for in-depth analysis. After the process is initiated, an automatic response is first executed on the edge side. The edge computing node that issued the alarm signal immediately initiates two preset operations. Operation one is to instantly boost the sampling frequency of the specific sensor channels (possibly one or more) related to this alarm from the normal basic monitoring frequency (e.g., 20Hz) to a preset high-frequency diagnostic frequency (e.g., 200Hz or 500Hz). Operation two is to simultaneously initiate a raw data buffer window with a duration of T1. The setting of T1 needs to be carefully considered to cover the critical period of abnormal development; for example, it can be set to 5 seconds before the alarm and 15 seconds after the alarm, for a total length of 20 seconds. During this window period, the "raw high-speed sampling data" (i.e., the data after frequency boosting) of the relevant sensors will be stored completely and without any compression in the local non-volatile memory (such as an SD card or Flash chip) of the edge node. This high-fidelity data is invaluable for subsequent detailed root cause analysis (such as identifying shocks and resonant frequencies).

[0070] Almost simultaneously, a response is also triggered on the cloud side. Upon receiving a lightweight alarm signal from the edge node, the cloud server immediately performs two operations. First, it marks the subsequent data stream corresponding to the alarm node as high priority in the data processing queue and may dynamically allocate dedicated computing resources, such as higher CPU cores, to ensure that the processing and analysis of the node's data can be performed preferentially and quickly. Second, the analysis module running on the cloud (possibly combining digital twins and deep learning models) quickly analyzes the initial alarm signal and the subsequently uploaded compressed data. If the analysis module initially determines that the anomaly is serious and the cause is complex, requiring more detailed raw waveform data for in-depth root cause analysis, the cloud server will automatically send a data retrieval command to the edge computing node that triggered the alarm.

[0071] Finally, the process enters the instruction execution phase. Upon receiving a clear data retrieval instruction from the cloud, the edge computing node located on-site will upload the previously cached local non-volatile memory (lasting T1 seconds) of raw, high-speed sampled data to the cloud server via an available network channel (possibly the link used for the alarm, or another backup link). This completes the entire collaborative process from low-latency alarm to high-fidelity data transmission in an anomaly event, and the cloud obtains the valuable raw data needed for in-depth analysis.

[0072] The collaborative response and data preservation process cleverly resolves the conflicting data requirements of real-time alarms and in-depth analysis (the former needs speed and simplicity, while the latter needs comprehensiveness and detail). By intelligently and conditionally caching high-fidelity raw data at the edge and retrieving it on demand in the cloud, it ensures the complete retention of critical data for abnormal events. This enables precise fault diagnosis, incident review, and liability determination afterward, introducing the "black box" concept into engineering monitoring.

[0073] In one specific implementation, a blockchain-based monitoring data storage and audit trail mechanism is also included, specifically: The cloud server generates hash digests with timestamps for key audit events generated during system operation in chronological order and packages them into data blocks; the key audit events include at least: original alarm signals reported by edge computing nodes, hierarchical early warning information generated in the cloud and its handling feedback status, phased records of the global confidence index of the digital twin model, important version update records of the incremental learning of the deep learning model, and execution records of data retrieval instructions; The data blocks are synchronized to a private blockchain network composed of nodes from the construction party, the supervision party, and the owner through a consensus mechanism for distributed storage. Any participating node can perform hash verification and tamper-proof traceability on the data blocks stored on the private blockchain network according to its permissions, so as to form a reliable construction safety audit log for the entire process.

[0074] In the above implementation, a blockchain-based monitoring data storage and audit trail mechanism is introduced to address the issues of easily questioned and difficult-to-verify authenticity of engineering monitoring data, thereby constructing a reliable security audit system throughout the entire process. The core operation of this mechanism is executed by a cloud server. The cloud server organizes and packages all key data and events generated during system operation. These key data and events specifically include, but are not limited to: raw alarm signals uploaded from edge nodes, tiered early warning information generated after cloud analysis, real-time updated global model confidence indicators from the digital twin model, records of incremental model learning by the deep learning model, and data retrieval instructions issued from the cloud to the edge. The server generates a digital fingerprint, or "hash value," with a precise timestamp for each event or batch of events according to the chronological order of their occurrence. These hash values ​​and related data index information are then packaged together to form an immutable data block.

[0075] Subsequently, these packaged data blocks are not simply stored in a cloud server database, but are synchronized to a specially constructed private blockchain network through a "consensus mechanism" (e.g., the Practical Byzantine Fault Tolerance (PBFT) algorithm). The participating nodes of this private chain are deployed and maintained by the construction party, the supervision party, and the owner, forming a multi-party distributed ledger. Each data block is synchronously recorded on the nodes of all participating parties, achieving distributed notarization. Once the data is recorded on the chain, due to its hash linkage and timestamp characteristics, no single party can tamper with the historical records. To tamper with such data would require controlling more nodes than the consensus mechanism requires, which is virtually impossible in a private blockchain with multi-party mutual supervision.

[0076] Finally, based on this evidence-based blockchain network, any participant (construction company, supervisor, owner) can access the data on the chain according to their permissions. They can perform hash verification on any data block to verify whether its content has been modified since it was recorded. At the same time, they can trace the entire chain of events from the start of monitoring to any given moment in time, making it impossible to tamper with the hash pointers between blocks. For example, "On a certain day, month, year, hour, and minute, edge node 3 reported a sudden anomaly; 10 seconds later, a level-two warning was generated in the cloud; 1 minute later, the supervisor node confirmed receipt of the warning; 5 minutes later, the on-site inspection results were reported and recorded on the chain." This forms a "full-process trusted construction safety audit log" that is witnessed by all parties and cannot be denied.

[0077] By introducing blockchain technology, an immutable, traceable, and mutually trusted data storage and auditing foundation is established for the entire intelligent monitoring and early warning system. This fundamentally eliminates the possibility of monitoring data being forged, altered, or deleted afterward, ensuring that every alarm record, every model update, and every handling instruction becomes irrefutable evidence.

[0078] In one specific implementation, the private blockchain network is equipped with smart contracts for automatically executing compliance procedures following an early warning response, specifically including: When the cloud server generates a Level 3 alert, the smart contract is automatically triggered; The smart contract executes the following logic: First, it forcibly pushes the warning information and the corresponding panoramic data storage block to all participating nodes and requests confirmation receipts; second, if no confirmation receipt is received from any key party node within a preset time, it automatically sends a compliance alert to the preset superior regulatory platform; finally, it records the entire timeline of this warning from its generation to the responses of all parties to the blockchain.

[0079] In the above implementation, smart contracts are further deployed to automate the compliance process following a high-level alert, ensuring that the alert response is not delayed or missed due to human factors. A smart contract is a pre-written piece of code deployed on the blockchain that executes automatically when certain conditions are met. In this solution, the smart contract is designed to listen for events where the cloud server generates alert information. Specifically, when the cloud server generates the highest-level "Level 3 alert" information, this event (or the data block containing the alert information being uploaded to the blockchain) will automatically trigger the associated smart contract to begin execution.

[0080] Once triggered, a smart contract will automatically execute according to its preset logic. Its execution logic typically includes several key steps. First, the contract will force the detailed information of this Level 3 warning (including the warning content and the hash of the associated data storage block) to be pushed through the blockchain network to all preset "participant nodes," namely the nodes of the construction party, supervisor, and owner, and simultaneously send a confirmation request instruction to these nodes. Second, the smart contract has a preset time window, such as 10 minutes or 30 minutes, which can be set according to the project's safety procedures. The contract will monitor and wait for confirmation responses from each key party node. If, within the preset time window, any party (especially the construction party or supervisor directly responsible for safety) fails to return a confirmation response, the smart contract will automatically execute the next step: sending a compliance alert to a preset higher-level regulatory platform (such as the construction unit's group safety monitoring center or the interface of the government safety supervision department), reporting when a Level 3 warning occurred in a certain project and that a responsible party failed to confirm its response within the specified time. Finally, regardless of the responses from all parties, the smart contract will record the entire timeline of the alert—from its generation and push to the responses (or failure to respond within the time limit)—onto the blockchain in a complete and immutable manner, forming a closed-loop audit trail.

[0081] By deploying smart contracts, the response process for high-level security alerts is transformed from reliance on potentially delayed or missed methods such as manual communication and telephone notifications into automated, mandatory, and non-repudiable procedural execution. Leveraging the rigidity of code, it ensures that the most urgent security information must reach all key responsible parties and be confirmed, and sets a response time limit. If the time limit is exceeded, the system automatically reports upwards, forming an effective supervision and accountability mechanism. This significantly strengthens the seriousness and enforceability of security alerts, avoids security response loopholes caused by human negligence or wishful thinking, and makes the entire security management system operate more standardized, efficient, and reliable.

[0082] To better understand the real-time stress monitoring and intelligent early warning method for hanging basket construction, the system for implementing this method is described below. The system includes: a multi-source sensor network deployed on the key load-bearing structure of the hanging basket, comprising at least strain sensors, tilt sensors, temperature sensors, and force sensors, used to simultaneously collect data on structural stress, deformation posture, ambient temperature, and key point reaction forces; a finite element digital twin model module deployed on a cloud server, used to dynamically simulate the mechanical state of the hanging basket under current construction loads and historical load sequences based on the hanging basket's design drawings, material properties, and real-time sensor data; and a stress anomaly identification model module deployed on a cloud server, whose inputs are pre-processed and feature-fused multi-source time-series sensor data and the real-time simulation output of the finite element digital twin model module, and whose outputs are a comprehensive stress state evaluation index and anomaly probability for key structural components. The cloud-edge collaborative computing architecture includes: at least one edge computing node deployed at the site of the hanging basket, used for real-time filtering, compression, and preliminary anomaly detection of data from the multi-source sensor network; a cloud server, used to receive data from the edge computing node, drive the finite element digital twin model module to perform simulation, and call the deep learning stress anomaly identification model module for in-depth analysis; if the comprehensive stress state evaluation index exceeds a preset threshold or the anomaly probability exceeds a warning value, a graded early warning information is generated; and a construction management terminal, used to receive and display the graded early warning information pushed by the cloud server.

[0083] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0084] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details.

Claims

1. A method for real-time stress monitoring and intelligent early warning during hanging basket construction, characterized in that, include: S1. A multi-type sensor network is deployed on the key load-bearing structure of the hanging basket. The sensor network includes at least strain sensors, tilt sensors, temperature sensors and force sensors, which are used to synchronously collect structural stress, deformation posture, ambient temperature data and key point reaction force data. S2. Establish a finite element digital twin model of the hanging basket. The finite element digital twin model dynamically simulates the mechanical state of the hanging basket under the current construction load and historical load sequence based on the design drawings, material properties and real-time sensor data of the hanging basket. S3. Construct a stress anomaly identification model based on deep learning. The input of the stress anomaly identification model is the real-time simulation output of multi-source time-series sensing data and digital twin model after preprocessing and feature fusion. The output is the comprehensive stress state evaluation index and anomaly probability of key structural parts. S4. Deploy a cloud-edge collaborative computing architecture, in which edge computing nodes are deployed at the site of the hanging basket to perform real-time filtering, compression and preliminary anomaly judgment on sensor data, and upload the processed data to the cloud server. After receiving the data, the cloud server drives the digital twin model to perform high-fidelity simulation and calls the stress anomaly identification model for in-depth analysis. If the comprehensive stress state assessment index exceeds the preset threshold or the anomaly probability exceeds the warning value, a graded early warning information is generated. The graded early warning information is pushed to the construction management terminal in real time. The graded early warning information includes at least the abnormal location, the degree of stress exceeding the limit, the possible failure mode, and the handling suggestions.

2. The method for real-time stress monitoring and intelligent early warning during hanging basket construction as described in claim 1, characterized in that, In step S1, the deployment of multiple sensor networks on the key load-bearing structure of the hanging basket specifically includes: The strain sensors are fixed by welding or bonding at the main truss of the hanging basket, the front / rear suspension points of the bottom basket, and the anchoring points of the traveling track to form multiple stress monitoring sections. The tilt sensor is installed at the cantilever end of the hanging basket and at the top node of the main truss; The temperature sensor is placed near the strain sensor to measure the local temperature of the structure; All sensors are protected by a waterproof and shockproof housing and are connected to the edge computing node via an industrial bus.

3. The method for real-time stress monitoring and intelligent early warning during hanging basket construction according to claim 1, characterized in that, In step S2, a finite element digital twin model of the hanging basket is established, specifically including: S21. Parametric Model Construction: Based on the design drawings of the hanging basket and the mechanical property parameters of the component materials, a parametric finite element digital twin model is generated in the cloud server by using a parametric script to drive the finite element analysis kernel. S22. Construction State Driven and Dynamic Load Mapping: Establish a digital state machine that is linked to the construction progress. Based on the received process transition signals, the digital state machine drives the finite element digital twin model to automatically switch to the corresponding predefined working condition mode and update the boundary conditions. In the concrete pouring process, the finite element digital twin model dynamically calculates and applies the distributed load of wet concrete based on the position information of the concrete placing boom. At the same time, it integrates the measured reaction force data from the force sensor at the bottom basket suspension point through parallel channels as a verification and supplementary input for the distributed load theory. During the traveling operation of the hanging basket, the finite element digital twin model automatically releases the rear anchor constraint and maps the jacking force and friction load in real time based on the data from the hydraulic system sensors. S23. Model Output Verification and Confidence Management: After each simulation calculation step, the finite element digital twin model compares the theoretical stress and strain values ​​of the key parts output by the model with the measured values ​​of the strain sensors at the corresponding locations in real time, and calculates and updates a global model confidence index accordingly.

4. The method for real-time stress monitoring and intelligent early warning during hanging basket construction as described in claim 3, characterized in that, Step S23 is executed through an adaptive multi-source data fusion engine, which specifically includes: a. Input preprocessing and credibility assessment: Calculate a dynamic credibility coefficient Ci(t) for each input data source; For simulation data from the finite element digital twin model, its confidence coefficient Cm(t) is directly assigned by the global model confidence index calculated in step S23; The reliability coefficient Csj(t) of the measured data from the j-th physical sensor is calculated by taking into account the variance of the sensor’s short-term historical data, the consistency with the readings of its spatially adjacent sensors, and its own health status score. b. Dynamic weight allocation and fusion calculation: For any key part k to be evaluated on the structure, its fusion weight Wik(t) at time t is dynamically allocated according to the rule Wik(t) = Ci(t) / Σ Ci(t) based on the credibility coefficient Ci(t) of each data source, and then the final fusion stress estimate Fk(t) = Σ[Wik(t) × Xik(t)] is calculated for this part, where Xik(t) is the stress estimate provided by the i-th data source; c. Anomaly Handling and Data Reconstruction: Continuously monitor the reliability coefficient Csj(t) and jump rate of each sensor data. When a sensor s data is determined to be abnormal, the data reconstruction mechanism is automatically triggered to increase the fusion weight of the simulation data at the corresponding location. Based on the data of spatially adjacent effective sensors, the supplementary estimate of the location is reconstructed through spatial interpolation algorithm and input into the fusion calculation.

5. The method for real-time stress monitoring and intelligent early warning during hanging basket construction as described in claim 1, characterized in that, In step S3, a deep learning-based stress anomaly recognition model is constructed, specifically including: S31. Model training phase: Supervised training of deep neural network is carried out using labeled multi-source sensor data from historical projects. The labels of multi-source sensor data include at least several working conditions such as normal, local overload, anchor loosening and structural cracking. The stress anomaly identification model composed of the deep neural network is trained to output anomaly type classification and overall anomaly probability. S32. Online monitoring and unknown pattern detection stage: The real-time data stream is input into the stress anomaly identification model for inference. At the same time, the deep feature representation inside the stress anomaly identification model is extracted, and the deep feature is compared online with the pre-stored normal working condition feature library. When the mathematical distance between the deep feature and all feature clusters in the normal feature library exceeds the dynamic threshold, and the confidence of the stress anomaly identification model in classifying the currently input known anomaly type is lower than the preset threshold, it is determined to be a potential unknown anomaly mode, and the corresponding data segment is stored in the suspicious mode library. S33. Model Incremental Evolution Stage: After the suspicious pattern library has accumulated to a preset size, new pattern samples confirmed and labeled by experts are merged with historical training data to form an augmented dataset. The importance weight of each parameter in the deep neural network of the stress anomaly identification model to the historical task is calculated, and constraints proportional to this importance are applied during the optimization process for the new data to carry out incremental training of the model, so as to achieve knowledge retention and iterative updates.

6. The method for real-time stress monitoring and intelligent early warning during hanging basket construction as described in claim 1 or 5, characterized in that, The tiered early warning information is divided into at least three levels, specifically: Level 1 warning is the attention level, which corresponds to the comprehensive stress state assessment index exceeding 80% of the theoretical reference value but not reaching the design value, or the probability of abnormality being in the range of 30%-60%. The handling recommendation is to strengthen manual inspection and data monitoring. Level 2 warning is a warning level, corresponding to the comprehensive stress state assessment index reaching or exceeding the design value but not reaching the safety factor tolerance value, or the abnormal probability being in the range of 60%-90%. The handling recommendation is to suspend the current loading or moving operation and conduct a comprehensive inspection. A Level 3 warning is an alarm level, corresponding to a comprehensive stress state assessment index exceeding the allowable value of the safety factor, or an abnormal probability exceeding 90%. The recommended response is to immediately organize personnel to evacuate the danger zone and activate the emergency rescue plan.

7. The method for real-time stress monitoring and intelligent early warning in hanging basket construction as described in claim 6, characterized in that, The cloud-edge collaborative computing architecture specifically includes a data preprocessing and fast response mechanism at the edge layer, which is implemented as follows: Edge hardware configuration and data interface: Embedded edge computing nodes are deployed at each key monitoring section of the hanging basket body. The edge computing nodes adopt low-power microprocessors and integrate multi-channel analog-to-digital converters and industrial fieldbus interfaces to directly connect to and poll the raw signals of strain sensors, tilt sensors and temperature sensors. Local rapid anomaly diagnosis rules: Each edge computing node runs a lightweight diagnostic program, which maintains a short-term historical data buffer based on a time window for each sensor channel. The lightweight diagnostic program calculates the deviation rate between the current sampled value of the sensor and the moving average of the buffer data for that channel in real time, using the formula: Deviation rate = |Current value - Moving average| / Moving standard deviation. When the deviation rate of any sensor channel continuously exceeds a preset first-level static threshold, it is determined as a potential slow anomaly. When the deviation rate instantaneously exceeds a preset second-level dynamic threshold, it is determined as a sudden and abrupt anomaly. For sudden and abrupt anomalies, the edge computing node immediately generates a lightweight alarm signal, which includes the abnormal sensor ID, anomaly type code, and a timestamp, and simultaneously triggers two independent communication links: the first link uploads data to the cloud server in real time via a low-power wide area network or a 5G slice network; the second link broadcasts data to the monitoring personnel's mobile terminal at the construction site via a local area network or Bluetooth. Data compression and asynchronous upload mechanism: In non-alarm state, the edge computing node performs lossy compression and downsampling preprocessing on the raw high-frequency sensor data collected; the rotating door trend compression algorithm is used to preserve data envelopment characteristics and reduce the sampling frequency from the original acquisition frequency to the engineering analysis frequency that satisfies the Nyquist theorem; the processed data is encapsulated into fixed-length data packets and a frame header, CRC checksum, and edge computing node ID are added; the data packets are uploaded to the cloud in batches through asynchronous transmission mode at preset fixed time periods or when the local cache reaches the capacity threshold.

8. The method for real-time stress monitoring and intelligent early warning during hanging basket construction according to claim 7, characterized in that, After the lightweight alarm signal is triggered, a collaborative response and data preservation process is executed, specifically as follows: Edge-side response: The edge computing node that issues a lightweight alarm signal immediately and automatically performs the following operations: a) Increases the sampling frequency of the sensor channel related to the alarm from the basic monitoring frequency to a preset high-frequency diagnostic frequency; b) Starts a raw data buffer window with a duration of T1, during which the raw high-speed sampling data of the relevant sensors is completely stored in the local non-volatile memory. Cloud-side response: After receiving a lightweight alarm signal, the cloud server performs the following operations: a) Marks the data stream of the edge computing node as high priority and allocates dedicated computing resources for processing; b. If its analysis module determines that high-fidelity data is needed for root cause analysis, it will automatically send a data retrieval command to the edge computing node. Instruction execution: After receiving the data retrieval instruction, the edge computing node uploads the cached raw high-speed sampling data to the cloud server.

9. The method for real-time stress monitoring and intelligent early warning during hanging basket construction as described in claim 8, characterized in that, It also includes a blockchain-based monitoring data storage and audit trail mechanism, specifically: The cloud server generates hash digests with timestamps for key audit events generated during system operation in chronological order and packages them into data blocks; the key audit events include at least: original alarm signals reported by edge computing nodes, hierarchical early warning information generated in the cloud and its handling feedback status, phased records of the global confidence index of the digital twin model, important version update records of the incremental learning of the deep learning model, and execution records of data retrieval instructions; The data blocks are synchronized to a private blockchain network composed of nodes from the construction party, the supervision party, and the owner through a consensus mechanism for distributed storage. Any participating node can perform hash verification and tamper-proof traceability on the data blocks stored on the private blockchain network according to its permissions, so as to form a reliable construction safety audit log for the entire process.

10. The method for real-time stress monitoring and intelligent early warning during hanging basket construction as described in claim 9, characterized in that, The private blockchain network is equipped with smart contracts for automatically executing compliance procedures following an alert response, specifically including: When the cloud server generates a Level 3 alert, the smart contract is automatically triggered; The smart contract executes the following logic: First, it forcibly pushes the warning information and the corresponding panoramic data storage block to all participating nodes and requests confirmation receipts; second, if no confirmation receipt is received from any key party node within a preset time, it automatically sends a compliance alert to the preset superior regulatory platform; finally, it records the entire timeline of this warning from its generation to the responses of all parties to the blockchain.

Images