A method for predicting risks during use of a scaffold

Abstract

The application discloses a kind of risk prediction methods in the process of scaffold, belong to scaffold technical field.It includes: in the key parts of scaffold deployment sensor cluster, real-time acquisition dynamic response and geometric deformation parameter;Physical information neural network is constructed, structural dynamics equation is embedded into loss function as physical constraint, network is trained and finite element model is corrected, obtain digital twin;By analyzing natural frequency variation, frequency band energy variation and modal curvature variation, overall stiffness degradation is judged, local damage is identified and quantified fastener loosening;Image is collected by camera, and construction conditions are identified and quantified by target detection algorithm;Damage information and working condition parameters are input into digital twin for stress evolution simulation, predict future stress response indicators and compare with safety threshold, generate graded warning information.Structure state and construction conditions are deeply integrated, with stress trend prediction capability, which improves risk prediction accuracy and timeliness.

Description

Technical Field

[0001] This invention belongs to the field of scaffolding technology, specifically relating to a risk prediction method during the use of scaffolding. Background Technology

[0002] Scaffolding, as a widely used temporary support structure in construction projects, directly impacts the safety of construction workers and the surrounding environment. Statistics show that scaffolding accidents account for a significant proportion of construction accidents, with overall instability and partial collapse caused by factors such as accumulated structural damage, excessive load variations, and loose fasteners being the main accident types. Therefore, real-time prediction and early warning of risks during scaffolding use have significant engineering value and practical implications.

[0003] Currently, the existing technologies for scaffolding safety monitoring and risk prediction mainly employ the following methods: Firstly, the scaffolding condition is assessed through regular manual inspections, using methods such as visual checks and random checks with torque wrenches. This method relies on personnel experience and suffers from problems such as strong subjectivity, limited inspection coverage, and inability to monitor continuously in real time, making it difficult to detect gradual damage and sudden load changes within the structure.

[0004] Secondly, structural analysis methods based on finite element modeling establish scaffolding models according to design parameters and perform static or dynamic analyses. These methods typically employ idealized boundary conditions and initial stiffness parameters, failing to reflect the actual state changes of the scaffolding during actual use due to factors such as loose fasteners, member deformation, and foundation settlement. A "digital gap" exists between the model and the actual structure, making it difficult to guarantee prediction accuracy.

[0005] Third, sensor-based early warning methods involve deploying sensors such as strain gauges and accelerometers at key parts of the scaffolding to collect real-time response data and setting a single threshold for exceeding limits. However, these methods typically only provide post-event warnings for abnormal responses that have already occurred, lacking the ability to predict the evolution of risks over future periods. Furthermore, the sensor data is not closely integrated with structural mechanics mechanisms, making it difficult to distinguish between different damage types and load sources, resulting in a high false alarm rate.

[0006] Fourth, machine vision-based work condition recognition methods identify construction loads such as personnel and materials through cameras. Existing methods mostly focus on the identification of static loads, lacking means to dynamically couple the identification results with the structural stress model. They cannot quantify the impact of load changes on structural safety, nor can they achieve linkage analysis between load distribution and structural response.

[0007] In summary, the existing technology has the following main shortcomings: It is impossible to achieve real-time dynamic integration of scaffolding structural status and construction conditions; The lack of methods to correct models by combining physical mechanisms with measured data leads to discrepancies between digital models and actual structures. It is difficult to predict the trend of force evolution in the future, and early warnings are mostly reactive. Without a closed-loop optimization mechanism, the prediction accuracy cannot be continuously improved as engineering data accumulates.

[0008] Therefore, there is an urgent need for a scaffolding risk prediction method that can integrate multi-source monitoring data, construct a high-fidelity digital twin model, predict stress evolution trends, and has self-optimization capabilities. Summary of the Invention

[0009] In view of this, the purpose of the present invention is to provide a risk prediction method during the use of scaffolding.

[0010] To achieve the above objectives, the present invention provides the following technical solution: A method for risk prediction during scaffolding use includes the following steps: S1. Deploy sensor clusters at key load-bearing members and connection nodes of the scaffolding to collect dynamic response parameters and geometric deformation parameters of the scaffolding in real time. S2. Based on the data collected in S1, a physical information neural network is constructed. The dynamic equation of the scaffold structure is embedded as a physical constraint term into the loss function of the neural network. The network is trained by minimizing the loss function containing the data fitting error and the physical equation residual. The stiffness parameters of the initial finite element model are corrected using the trained network parameters to obtain a digital twin that matches the actual state of the scaffold. S3. Using the digital twin obtained in S2, combined with the data collected in real time in S1, the overall stiffness degradation of the frame is judged by analyzing the change of the natural frequency of the members, the location of local damage is identified by analyzing the change of vibration signal frequency band energy, and the loosening of fastener nodes is located and the degree of loosening is quantified by calculating the change of modal curvature. S4. Images are collected by cameras deployed at the construction site, and the construction conditions are identified by target detection algorithms. The construction conditions include personnel operations and material stacking. The conditions are quantified based on the identification results, including personnel distribution density, personnel concentration areas, stacking location and stacking amount. S5. Using the damage information identified in S3 and the working condition parameters quantified in S4 as boundary conditions, input the digital twin of S2 to simulate the structural stress evolution and predict the stress response index of the frame in the future period. The stress response index includes member stress, frame deformation and overall stability coefficient. S6. Compare the force response index predicted by S5 with the preset safety threshold. When the predicted value exceeds the set threshold, generate graded early warning information and push it through on-site and remote means. The early warning information includes the location of the risk, the type of risk, the expected time of occurrence, and disposal suggestions.

[0011] As a further preferred embodiment of the present invention, it also includes S7, comparing the actual risk event with the prediction result of S5, calculating the prediction error, and using the prediction error to perform incremental learning optimization on the digital twin in S2.

[0012] As a further preferred embodiment of the present invention, the key load-bearing members in S1 include the bottom uprights, the ground-level bracing, the scissor bracing members, and the uprights at the corners; the connection nodes include the fastener connection nodes between the uprights and the horizontal bars and the fastener connection nodes between the scissor bracing and the uprights; the sensor cluster includes axial force gauges deployed on the bottom uprights and the ground-level bracing, accelerometers arranged along the height direction of the frame, inclinometers deployed on the top and bottom of the frame, and strain gauges deployed on the surfaces of the ground-level bracing and the scissor bracing members; the dynamic response parameters include the vibration acceleration of the members, the natural frequency of the members, the mode shape, and the damping ratio, and the geometric deformation parameters include the verticality of the frame, the foundation settlement, the horizontal displacement of the top of the frame, and the axial strain of the members.

[0013] As a further preferred embodiment of the present invention, S2 specifically includes: S21. Establish an initial finite element model based on the upright spacing, step distance, ground brace height, scissor brace arrangement, and wall tie spacing in the scaffolding construction plan. S22. Construct a physical information neural network. The input layer of the neural network includes time coordinates and spatial coordinates. The output layer includes three-dimensional displacement, velocity and acceleration. The hidden layer adopts a 5-layer fully connected structure with 100 nodes in each layer. The activation function is the tanh function. S23. Constructing the loss function ,in, For balance coefficient, The mean square error between the network output and the measured data of S1 is calculated using the following formula: Where N is the sample size. , and These are the displacement, velocity, and acceleration data of the i-th sample point measured in S1, respectively. , and These are the displacement, velocity, and acceleration data of the i-th sample point predicted by the neural network; The mean square value of the residual obtained after substituting the displacement field output by the network into the dynamic equation of the scaffold structure is calculated using the following formula: Where M is the number of physical constraint sampling points; The mass matrix of the scaffolding structure. Here is the damping matrix of the scaffolding structure. Here is the stiffness matrix of the scaffolding structure. The time corresponding to the j-th sampling point The external load vector, Let j be the displacement vector of the j-th sampling point predicted by the neural network. Let be the velocity vector of the j-th sampling point predicted by the neural network. Let be the acceleration vector of the j-th sampling point predicted by the neural network; S24. The Adam optimizer is used to train the neural network, and the network parameters are optimized by minimizing the loss function; S25. Use the trained network parameters to correct the stiffness parameters of the initial finite element model to obtain a digital twin that matches the actual state of the frame.

[0014] As a further preferred embodiment of the present invention, S3 specifically includes: S31. Perform a fast Fourier transform on the acceleration data collected by S1, extract the first 5 natural frequencies, calculate the relative rate of change between the measured frequency and the reference frequency output by the digital twin of S2, and determine the overall stiffness degradation of the frame when the decrease of any one frequency exceeds 5%. S32. Perform wavelet packet decomposition on the vibration signal of each acceleration measuring point, extract the energy distribution of each frequency band, calculate the rate of change between the measured energy and the reference energy, and determine that there is local damage to the rod near the measuring point when the rate of change of energy at a certain measuring point in the 8~16Hz frequency band exceeds 20%. S33. Based on the extracted mode shape, calculate the modal curvature at each node position using the central difference method. Calculate the relative change between the measured modal curvature and the reference modal curvature. When the relative change exceeds 10%, it is determined that the fastener at that node is loose. Based on the pre-calibrated fastener torque-modal curvature change relationship curve, the missing value of the fastener torque is inverted to quantify the degree of looseness.

[0015] As a further preferred embodiment of the present invention, S4 specifically includes: S41. Deploy network cameras around the scaffolding work area and transmit video streams to the AI ​​analysis server in real time via network cable or 5G network; S42, the AI ​​analysis server decodes and preprocesses the video stream, performing frame extraction at a frequency of 2 to 5 frames per second. S43. The YOLOv8 target detection model is used to perform real-time analysis of the frame-sampling images to identify personnel, steel pipe bundles, sandbags and fastener boxes, and the DeepSORT algorithm is used to continuously track the identified targets. S44. Based on the target detection and tracking results, the scaffolding working surface is divided into grid areas. The number of people in each grid area is counted, and the personnel distribution density is calculated. When the personnel density of a continuous area exceeds 1 person / m², it is determined to be a concentrated area of ​​personnel. S45. By identifying the number of steel pipe bundles and sandbags, and combining the preset single-piece weight, the load capacity is calculated, and the load position is determined by detecting the center coordinates of the target frame. S46. Organize the personnel distribution density, personnel concentration areas, loading location and loading capacity into structured data, and use it as the boundary condition input for S5.

[0016] As a further preferred embodiment of the present invention, S5 specifically includes: S51. The damage information identified in S3 is converted into stiffness reduction parameters of the digital twin, wherein: the overall stiffness degradation is proportionally reduced to the overall stiffness matrix of the digital twin based on the relative change rate of the natural frequency calculated in S31; the local damage of the members is reduced according to the location and degree of damage identified in S32, and the elastic modulus of the damaged member element is reduced accordingly; the fastener loosening is reduced according to the loose node and torque missing value identified in S33, and the rotational stiffness of the node is reduced accordingly. S52. The working condition parameters obtained by S4 quantization are transformed into time-varying load boundary conditions of the digital twin, wherein: personnel distribution density and personnel concentration area are transformed into uniformly distributed live load of the corresponding area; load location and load amount are transformed into concentrated load applied to the corresponding node. S53. Set the prediction time window and time step. Use the digital twin updated in S51 and the time-varying load generated in S52 as the initial state and external excitation. Use the step-by-step integration method to perform structural dynamic time history analysis and solve the stress of the frame members, the deformation of the frame and the overall stability coefficient in the future period. S54 outputs the maximum stress value of each key stress member, the maximum horizontal displacement value of the top of the frame, and the overall stability coefficient, forming the time series prediction results of the structural stress response.

[0017] As a further preferred embodiment of the present invention, S6 specifically includes: S61. Preset safety thresholds for member stress, frame deformation, and overall stability coefficient; S62. Compare the time series prediction results output by S54 with the preset safety threshold. When any prediction indicator exceeds the corresponding safety threshold, classify the warning level according to the degree of exceeding the limit. S63. Based on the prediction results output by S54, combined with the damage location identified by S3 and the concentrated working condition area identified by S4, determine the location of the risk occurrence; determine the risk type based on the type of index exceeding the limit. S64. Estimate the timing of risk occurrence based on current operating condition trends and provide corresponding handling recommendations; S65. Package the location of the risk, the type of risk, the expected time of occurrence, the handling recommendations, and the warning level into a warning message and push it out through on-site and remote means.

[0018] As a further preferred embodiment of the present invention, S7 specifically includes: S71. Compare the actual risk events that occur after the warning is pushed out by S6 with the force response indicators predicted by S5, and record the actual occurrence time, location, type and severity of the risk events. S72. Calculate the prediction error based on the comparison results, including time deviation, location deviation and index deviation; S73. Construct a loss function using the prediction error, perform incremental learning on the digital twin in S2, update the network parameters of the physical information neural network, and optimize the stiffness parameters and damage recognition accuracy of the digital twin. S74. Use the updated digital twin as the initial model for subsequent predictions to achieve continuous self-optimization of the method.

[0019] The beneficial effects of this invention are as follows: 1. Real-time dynamic fusion of structural state and construction conditions: This invention deploys a sensor cluster to collect dynamic response parameters and geometric deformation parameters of the frame in real time. Simultaneously, it employs a target detection algorithm to identify construction conditions (personnel operations, material loading) in real time. The identified damage information and quantified condition parameters are used together as boundary condition inputs for the digital twin. Compared to the existing approach of separating sensor monitoring and condition identification, this invention achieves deep fusion of structural response data and construction load information, making risk prediction more closely aligned with actual construction scenarios.

[0020] 2. A high-fidelity digital twin combining physical mechanisms and measured data was constructed: This invention constructs a physical information neural network, embedding the scaffolding structure dynamic equations as physical constraints into the loss function. The network is trained by minimizing the combined loss function of data fitting error and physical equation residuals, and the stiffness parameters of the initial finite element model are corrected using the trained network parameters. Compared to the problem of model-solid disconnect in existing finite element modeling methods, this invention organically integrates measured data with physical mechanisms, obtaining a high-fidelity digital twin that matches the actual state of the scaffolding, significantly improving model accuracy.

[0021] 3. This invention achieves the ability to predict future stress evolution trends: It inputs damage information and operating parameters as boundary conditions into a digital twin, and predicts the stress on the frame members, frame deformation, and overall stability coefficient in future periods through structural stress evolution simulation. The predicted results are compared with safety thresholds to generate tiered early warning information. Compared to the limitations of existing technologies that can only provide post-event over-limit alarms, this invention achieves forward-looking prediction of risk evolution trends, buying valuable emergency response time for construction personnel.

[0022] 4. A closed-loop self-optimization mechanism based on incremental learning is established: This invention compares the actual risk events with the prediction results, calculates the prediction error, uses the error to incrementally learn and optimize the digital twin, and uses the optimized model as the initial model for subsequent predictions. Compared with the shortcomings of existing technologies that cannot continuously improve accuracy with the accumulation of engineering data, this invention forms a closed-loop mechanism of "prediction-verification-optimization," enabling the prediction accuracy to continuously improve with the increase of data used.

[0023] 5. A multi-scale, multi-dimensional damage identification system was constructed: This invention determines the overall stiffness degradation of the frame by analyzing the changes in the natural frequencies of the members, identifies the location of local damage by analyzing the changes in the frequency band energy of vibration signals, and locates and quantifies the degree of loosening of fastener nodes by calculating the changes in modal curvature. Compared with the problem that existing technologies have difficulty distinguishing different types of damage, this invention forms a multi-scale damage identification system from the overall to the local, and from qualitative to quantitative, providing more accurate input for risk warning.

[0024] In summary, this invention effectively solves the technical problems of large model deviation, unpredictability, difficulty in fusion, and inability to self-optimize in the prior art by integrating multi-source monitoring data, constructing a high-fidelity digital twin model, realizing the prediction of stress evolution trends, and establishing a closed-loop self-optimization mechanism. It significantly improves the accuracy of risk prediction and the timeliness of early warning during the use of scaffolding. Attached Figure Description

[0025] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram of the process of S2 in this invention; Figure 3 This is a schematic diagram of the process of S3 in this invention. Detailed Implementation

[0026] like Figures 1-3 As shown, the present invention discloses a risk prediction method in the use of scaffolding, including the following steps S1 to S7.

[0027] S1: Sensor Deployment and Data Acquisition This step involves deploying a cluster of sensors at key load-bearing members and connection nodes of the scaffolding to collect dynamic response parameters and geometric deformation parameters of the scaffolding in real time.

[0028] Specifically, the key load-bearing members include the bottom uprights, the base braces, the scissor braces, and the uprights at corners; the connection nodes include the fastener connections between the uprights and horizontal bars, and the fastener connections between the scissor braces and the uprights. The sensor cluster includes: axial force gauges deployed on the bottom uprights and base braces to measure the axial force of the members; accelerometers arranged along the height of the frame to collect vibration signals; inclinometers deployed at the top and bottom of the frame to measure the verticality of the frame and the horizontal displacement at the top; and strain gauges deployed on the surfaces of the base braces and scissor braces to measure the strain of the members.

[0029] Dynamic response parameters include member vibration acceleration, member natural frequency, mode shape, and damping ratio; geometric deformation parameters include frame verticality, foundation settlement, horizontal displacement at the top of the frame, and member axial strain.

[0030] By deploying multiple types of sensors at key locations, multi-dimensional and real-time monitoring of the scaffolding structure's condition can be achieved, providing an accurate data foundation for the subsequent construction of digital twins and damage identification.

[0031] S2: Digital Twin Construction Based on the data collected by S1, this step constructs a Physics-Informed Neural Network (PINN). The scaffolding structure dynamics equations are embedded as physical constraints into the loss function of the neural network. The network is trained by minimizing the loss function, which includes data fitting errors and physical equation residuals. The stiffness parameters of the initial finite element model are corrected using the trained network parameters to obtain a digital twin that matches the actual state of the scaffolding.

[0032] Specifically, it includes the following sub-steps: S21: Establish the initial finite element model An initial finite element model was established based on the upright spacing, step distance, ground beam height, scissor bracing arrangement, and wall tie spacing specified in the scaffolding construction plan. This model reflects the design state of the scaffolding.

[0033] S22: Constructing a physical information neural network The input layer of the neural network includes time coordinates. and spatial coordinates The output layer includes three-dimensional displacement, velocity, and acceleration data. The hidden layer uses a 5-layer fully connected structure, with 100 nodes in each layer, and the activation function is... function.

[0034] S23: Constructing the loss function loss function Data fitting term and physical constraints composition: in, This is a balancing coefficient used to adjust the weight between the two items, ranging from 0.1 to 0.5.

[0035] Data fitting term The mean square error between the network output and the measured data of S1 is calculated using the following formula: in, For the sample size, , and The actual measured number in S1 Displacement, velocity, and acceleration data for each sample point. , and These are the predictions of the neural network for the first time. Displacement, velocity, and acceleration data for each sample point.

[0036] Physical constraints The mean square value of the residual obtained after substituting the displacement field output by the network into the dynamic equation of the scaffold structure is calculated using the following formula: in, The number of physical constraint sampling points; The mass matrix of the scaffolding structure. Here is the damping matrix. Here is the stiffness matrix; For the first Each sampling point corresponds to a time The external load vector; , and These are the predictions of the neural network for the first time. The displacement, velocity, and acceleration vectors of each sampling point.

[0037] S24: Network Training The Adam optimizer is used to train the neural network by minimizing the loss function. Optimize network parameters. The Adam optimizer can adaptively adjust the learning rate, accelerating convergence.

[0038] S25: Model Correction The stiffness parameters of the initial finite element model are corrected using the trained network parameters to obtain a digital twin that matches the actual state of the frame.

[0039] By embedding the structural dynamics equations as physical constraints into the loss function through a physical information neural network, the network can follow physical laws while fitting measured data, thus solving the problem of the disconnect between the traditional finite element model and the actual entity and obtaining a high-fidelity digital twin.

[0040] S3: Damage Identification This step utilizes the digital twin obtained in S2, combined with the real-time data collected in S1, to determine the overall stiffness degradation of the frame by analyzing the changes in the natural frequency of the members, to identify the location of local damage by analyzing the changes in the frequency band energy of the vibration signal, and to locate and quantify the loosening of fastener nodes by calculating the changes in modal curvature.

[0041] Specifically, it includes the following sub-steps: S31: Overall Stiffness Degradation Judgment Perform a Fast Fourier Transform (FFT) on the acceleration data acquired by S1 to extract the first 5 natural frequencies. Calculate the measured frequency and the S2 digital twin output reference frequency. Relative rate of change: When the decrease in any first-order frequency exceeds 5%, the overall stiffness of the frame is determined to be degraded.

[0042] S32: Local Damage Identification Wavelet packet decomposition was performed on the vibration signal at each acceleration measurement point to extract the energy distribution of each frequency band. Calculate the measured energy and the reference energy. Rate of change: When the energy change rate at a certain measuring point exceeds 20% in the 8-16Hz frequency band, it is determined that there is local damage to the rod near the measuring point.

[0043] S33: Fastener Loosening Positioning and Quantification Based on the extracted modal shapes The modal curvature at each node position is calculated using the central difference method: Calculate the measured modal curvature With reference modal curvature The relative change: When the relative change exceeds 10%, the fastener at that node is determined to be loose, and the missing value of the fastener torque is calculated by reversing the pre-calibrated fastener torque-modal curvature change relationship curve. Quantify the degree of loosening.

[0044] A multi-scale damage identification system was constructed, ranging from overall to local and from qualitative to quantitative analysis. This system can simultaneously identify overall stiffness degradation, local member damage, and fastener loosening, providing accurate damage information input for risk prediction.

[0045] S4: Operating Condition Identification and Quantization This step involves capturing images using cameras deployed at the construction site, employing object detection algorithms to identify construction conditions, including personnel operations and material loading, and quantifying the condition parameters based on the identification results.

[0046] Specifically, it includes the following sub-steps: S41: Video Acquisition and Transmission Network cameras are deployed around the scaffolding work area, and video streams are transmitted in real time to the AI ​​analysis server via Ethernet cable or 5G network.

[0047] S42: Image Preprocessing The AI ​​analysis server decodes and preprocesses the video stream, performing frame extraction at a frequency of 2-5 frames per second to reduce computational load.

[0048] S43: Target Detection and Tracking The YOLOv8 target detection model is used to analyze the frame-by-frame images in real time to identify people, steel pipe bundles, sandbags, and fastener boxes, and the DeepSORT algorithm is used to continuously track the identified targets to maintain the continuity of the target's identity.

[0049] S44: Quantification of Personnel Distribution Based on the target detection and tracking results, the scaffolding working surface is divided into... The grid area is used to count the number of people in each grid area. Calculate the population distribution density: in The area represents the grid area. When the population density in a continuous area exceeds 1 person / m², it is considered a densely populated area.

[0050] S45: Load Quantization By identifying the number of steel pipe bundles and sandbags and Combined with the preset single-piece weight and Calculate the load capacity: And by detecting the center coordinates of the target bounding box Determine the loading location.

[0051] S46: Data Output population distribution density Coordinates of the personnel concentration area, loading location and loading capacity Organize the data into structured form and use it as the boundary condition input for S5.

[0052] By using machine vision technology to identify construction conditions and quantify load parameters in real time, dynamic coupling of condition information and structural stress model is achieved, providing accurate load boundary conditions for risk prediction.

[0053] S5: Stress Evolution Simulation and Prediction This step uses the damage information identified in S3 and the working condition parameters quantified in S4 as boundary conditions, inputs them into the digital twin of S2 to simulate the structural stress evolution, and predicts the stress response index of the frame in the future period.

[0054] Specifically, it includes the following sub-steps: S51: Damage Information Conversion The damage information identified by S3 is converted into stiffness reduction parameters of the digital twin: Overall stiffness degradation: the relative rate of change of natural frequency calculated based on S31. The overall stiffness matrix of the digital twin is reduced proportionally: in This is the reduction factor; Localized damage to members: Based on the location and extent of damage identified by S32, the elastic modulus of the damaged member element is reduced accordingly. : in This is the reduction factor; Loose fasteners: Based on the loosening points identified by S33 and the missing torque value This corresponds to reducing the rotational stiffness of the nodes. : in These are calibration coefficients.

[0055] S52: Operating Parameter Conversion The load parameters obtained from S4 quantization are transformed into time-varying load boundary conditions for a digital twin: Population density and areas of high population concentration are converted into uniformly distributed live loads for the corresponding areas. ; Stacking location and load capacity Transformed into concentrated loads applied to the corresponding nodes. .

[0056] S53: Dynamic Time History Analysis Setting the prediction time window (e.g., 5-30 minutes) and time steps (e.g., 0.1 seconds), using the digital twin updated by S51 and the time-varying load generated by S52 as the initial state and external excitation, and employing Newmark- Structural dynamic time history analysis using the successive integration method. (Newmark-) The basic recursive formula for the law is: Solve for the stress of the frame members in the future time period. Deformation of the frame and overall stability coefficient .

[0057] S54: Output Results Output the maximum stress value of each key load-bearing member. Maximum horizontal displacement value at the top of the frame and overall stability coefficient This generates time-series prediction results of the structural stress response.

[0058] By inputting damage information and operating parameters as boundary conditions into a high-fidelity digital twin, and using dynamic time history analysis, quantitative prediction of future stress trends can be achieved, providing a scientific basis for forward-looking early warning.

[0059] S6: Tiered Early Warning This step compares the force response index predicted by S5 with the preset safety threshold. When the predicted value exceeds the set threshold, a graded early warning message is generated and pushed out through on-site and remote means.

[0060] Specifically, it includes the following sub-steps: S61: Safety threshold preset Based on scaffolding design specifications and engineering experience, a pre-set safety threshold for stress in the structural members is established. Safety threshold for frame deformation and overall stability coefficient safety threshold .

[0061] S62: Comparison and Classification The time series prediction results output by S54 are compared hourly with preset safety thresholds. An alert is triggered when any predicted indicator exceeds the corresponding safety threshold. The severity of the exceedance is determined by the degree of expiration. Classification of warning levels: Predicted value safety threshold Exceeding the limit by 0-10% triggers a Level 3 warning, 10%-30% triggers a Level 2 warning, and exceeding 30% triggers a Level 1 warning.

[0062] S63: Risk Identification and Type Assessment Based on the prediction results output by S54, combined with the damage location identified by S3 and the concentrated working condition area identified by S4, the location of the risk is determined; the risk type is determined according to the type of index exceeding the limit: exceeding the limit of member stress corresponds to member strength risk, exceeding the limit of frame deformation corresponds to overall stability risk, and exceeding the limit of overall stability coefficient corresponds to instability risk.

[0063] S64: Matching Time Estimates with Recommendations Based on the current trend of changes in operating conditions, the time of risk occurrence is estimated, and the corresponding handling suggestions are matched by calling the preset risk handling knowledge base.

[0064] S65: Information Push The location, type, expected time of occurrence, handling recommendations, and warning level of the risk are packaged into a warning message and simultaneously pushed through on-site audible and visual alarms, large screens at the construction site, and remote management platforms.

[0065] A tiered early warning mechanism has been implemented, with differentiated response measures taken based on the severity of the risk, and early warning information is promptly delivered to relevant personnel through multiple channels.

[0066] S7: Incremental Learning Self-Optimization This step compares the actual risk events with the prediction results of S5, calculates the prediction error, and uses the prediction error to perform incremental learning and optimization on the digital twin in S2.

[0067] Specifically, it includes the following sub-steps: S71: Event Comparison The actual occurrence of risk events after the S6 warning is pushed out is compared with the force response indicators predicted by S5, and the actual occurrence time of the risk events is recorded. ,Location ,type and severity .

[0068] S72: Error Calculation The prediction error, including time deviation, is calculated based on the comparison results. Positional deviation and indicator deviation .

[0069] S73: Incremental Learning Constructing a loss function using prediction error Incremental learning is performed on the digital twin in S2 to update the network parameters of the physical information neural network, thereby optimizing the stiffness parameters and damage recognition accuracy of the digital twin. in The number of risk events. For predicted values, These are actual observed values.

[0070] S74: Model Update The updated digital twin is used as the initial model for subsequent predictions, enabling continuous self-optimization of the method.

[0071] A closed-loop mechanism of "prediction-verification-optimization" has been established, which enables the prediction accuracy to continuously improve with the accumulation of engineering data and realize the continuous evolution of the model.

Claims

1. A risk prediction method during the use of scaffolding, characterized in that, Includes the following steps: S1. Deploy sensor clusters at key load-bearing members and connection nodes of the scaffolding to collect dynamic response parameters and geometric deformation parameters of the scaffolding in real time. S2. Based on the data collected in S1, a physical information neural network is constructed. The dynamic equation of the scaffold structure is embedded as a physical constraint term into the loss function of the neural network. The network is trained by minimizing the loss function containing the data fitting error and the physical equation residual. The stiffness parameters of the initial finite element model are corrected using the trained network parameters to obtain a digital twin that matches the actual state of the scaffold. S3. Using the digital twin obtained in S2, combined with the data collected in real time in S1, the overall stiffness degradation of the frame is judged by analyzing the change of the natural frequency of the members, the location of local damage is identified by analyzing the change of vibration signal frequency band energy, and the loosening of fastener nodes is located and the degree of loosening is quantified by calculating the change of modal curvature. S4. Images are collected by cameras deployed at the construction site, and the construction conditions are identified by target detection algorithms. The construction conditions include personnel operations and material stacking. The conditions are quantified based on the identification results, including personnel distribution density, personnel concentration areas, stacking location and stacking amount. S5. Using the damage information identified in S3 and the working condition parameters quantified in S4 as boundary conditions, input the digital twin of S2 to simulate the structural stress evolution and predict the stress response index of the frame in the future period. The stress response index includes member stress, frame deformation and overall stability coefficient. S6. Compare the force response index predicted by S5 with the preset safety threshold. When the predicted value exceeds the set threshold, generate graded early warning information and push it through on-site and remote means. The early warning information includes the location of the risk, the type of risk, the expected time of occurrence, and disposal suggestions.

2. The risk prediction method during the use of scaffolding according to claim 1, characterized in that: It also includes S7, which compares the actual risk events with the prediction results of S5, calculates the prediction error, and uses the prediction error to perform incremental learning and optimization on the digital twin in S2.

3. The risk prediction method during the use of scaffolding according to claim 1, characterized in that: Key load-bearing members in S1 include the bottom uprights, ground braces, scissor braces, and uprights at corners; connection nodes include fastener connections between uprights and horizontal bars and between scissor braces and uprights; the sensor cluster includes axial force gauges deployed on the bottom uprights and ground braces, accelerometers arranged along the height of the frame, inclinometers deployed on the top and bottom of the frame, and strain gauges deployed on the surfaces of the ground braces and scissor braces; dynamic response parameters include member vibration acceleration, member natural frequency, mode shape, and damping ratio; geometric deformation parameters include frame verticality, foundation settlement, horizontal displacement at the top of the frame, and axial strain of the members.

4. The risk prediction method during the use of scaffolding according to claim 1, characterized in that: S2 specifically includes: S21. Establish an initial finite element model based on the upright spacing, step distance, ground brace height, scissor brace arrangement, and wall tie spacing in the scaffolding construction plan. S22. Construct a physical information neural network. The input layer of the neural network includes time coordinates and spatial coordinates. The output layer includes three-dimensional displacement, velocity and acceleration. The hidden layer adopts a 5-layer fully connected structure with 100 nodes in each layer. The activation function is the tanh function. S23. Constructing the loss function ,in, For balance coefficient, The mean square error between the network output and the measured data of S1 is calculated using the following formula: Where N is the sample size. , and These are the displacement, velocity, and acceleration data of the i-th sample point measured in S1, respectively. , and These are the displacement, velocity, and acceleration data of the i-th sample point predicted by the neural network; The mean square value of the residual obtained after substituting the displacement field output by the network into the dynamic equation of the scaffold structure is calculated using the following formula: Where M is the number of physical constraint sampling points; The mass matrix of the scaffolding structure. Here is the damping matrix of the scaffolding structure. Here is the stiffness matrix of the scaffolding structure. The time corresponding to the j-th sampling point The external load vector, Let j be the displacement vector of the j-th sampling point predicted by the neural network. Let be the velocity vector of the j-th sampling point predicted by the neural network. Let be the acceleration vector of the j-th sampling point predicted by the neural network; S24. The Adam optimizer is used to train the neural network, and the network parameters are optimized by minimizing the loss function; S25. Use the trained network parameters to correct the stiffness parameters of the initial finite element model to obtain a digital twin that matches the actual state of the frame.

5. The risk prediction method during the use of scaffolding according to claim 1, characterized in that: S3 specifically includes: S31. Perform a fast Fourier transform on the acceleration data collected by S1, extract the first 5 natural frequencies, calculate the relative rate of change between the measured frequency and the reference frequency output by the digital twin of S2, and determine the overall stiffness degradation of the frame when the decrease of any one frequency exceeds 5%. S32. Perform wavelet packet decomposition on the vibration signal of each acceleration measuring point, extract the energy distribution of each frequency band, calculate the rate of change between the measured energy and the reference energy, and determine that there is local damage to the rod near the measuring point when the rate of change of energy at a certain measuring point in the 8~16Hz frequency band exceeds 20%. S33. Based on the extracted mode shape, calculate the modal curvature at each node position using the central difference method. Calculate the relative change between the measured modal curvature and the reference modal curvature. When the relative change exceeds 10%, it is determined that the fastener at that node is loose. Based on the pre-calibrated fastener torque-modal curvature change relationship curve, the missing value of the fastener torque is inverted to quantify the degree of looseness.

6. The risk prediction method during the use of scaffolding according to claim 5, characterized in that: S4 specifically includes: S41. Deploy network cameras around the scaffolding work area and transmit video streams to the AI ​​analysis server in real time via network cable or 5G network; S42, the AI ​​analysis server decodes and preprocesses the video stream, performing frame extraction at a frequency of 2 to 5 frames per second. S43. The YOLOv8 target detection model is used to perform real-time analysis of the frame-sampling images to identify personnel, steel pipe bundles, sandbags and fastener boxes, and the DeepSORT algorithm is used to continuously track the identified targets. S44. Based on the target detection and tracking results, the scaffolding working surface is divided into grid areas. The number of people in each grid area is counted, and the personnel distribution density is calculated. When the personnel density of a continuous area exceeds 1 person / m², it is determined to be a concentrated area of ​​personnel. S45. By identifying the number of steel pipe bundles and sandbags, and combining the preset single-piece weight, the load capacity is calculated, and the load position is determined by detecting the center coordinates of the target frame. S46. Organize the personnel distribution density, personnel concentration areas, loading location and loading capacity into structured data, and use it as the boundary condition input for S5.

7. The risk prediction method during the use of scaffolding according to claim 6, characterized in that: S5 specifically includes: S51. The damage information identified in S3 is converted into stiffness reduction parameters of the digital twin, wherein: the overall stiffness degradation is proportionally reduced to the overall stiffness matrix of the digital twin based on the relative change rate of the natural frequency calculated in S31; the local damage of the members is reduced according to the location and degree of damage identified in S32, and the elastic modulus of the damaged member element is reduced accordingly; the fastener loosening is reduced according to the loose node and torque missing value identified in S33, and the rotational stiffness of the node is reduced accordingly. S52. The working condition parameters obtained by S4 quantization are transformed into time-varying load boundary conditions of the digital twin, wherein: personnel distribution density and personnel concentration area are transformed into uniformly distributed live load of the corresponding area; load location and load amount are transformed into concentrated load applied to the corresponding node. S53. Set the prediction time window and time step. Use the digital twin updated in S51 and the time-varying load generated in S52 as the initial state and external excitation. Use the step-by-step integration method to perform structural dynamic time history analysis and solve the stress of the frame members, the deformation of the frame and the overall stability coefficient in the future period. S54 outputs the maximum stress value of each key stress member, the maximum horizontal displacement value of the top of the frame, and the overall stability coefficient, forming the time series prediction results of the structural stress response.

8. The risk prediction method during the use of scaffolding according to claim 7, characterized in that: S6 specifically includes: S61. Preset safety thresholds for member stress, frame deformation, and overall stability coefficient; S62. Compare the time series prediction results output by S54 with the preset safety threshold. When any prediction indicator exceeds the corresponding safety threshold, classify the warning level according to the degree of exceeding the limit. S63. Based on the prediction results output by S54, combined with the damage location identified by S3 and the concentrated working condition area identified by S4, determine the location of the risk occurrence; determine the risk type based on the type of index exceeding the limit. S64. Estimate the timing of risk occurrence based on current operating condition trends and provide corresponding handling recommendations; S65. Package the location of the risk, the type of risk, the expected time of occurrence, the handling recommendations, and the warning level into a warning message and push it out through on-site and remote means.

9. The risk prediction method during the use of scaffolding according to claim 2, characterized in that: S7 specifically includes: S71. Compare the actual risk events that occur after the warning is pushed out by S6 with the force response indicators predicted by S5, and record the actual occurrence time, location, type and severity of the risk events. S72. Calculate the prediction error based on the comparison results, including time deviation, location deviation and index deviation; S73. Construct a loss function using the prediction error, perform incremental learning on the digital twin in S2, update the network parameters of the physical information neural network, and optimize the stiffness parameters and damage recognition accuracy of the digital twin. S74. Use the updated digital twin as the initial model for subsequent predictions to achieve continuous self-optimization of the method.

Images