A construction hanging basket operation risk real-time monitoring method and system

Abstract

The application provides a construction hanging basket operation risk real-time monitoring method and system, and belongs to the technical field of monitoring and early warning. The method comprises the following steps: analyzing acquired video monitoring data based on computer vision technology, and extracting working condition information and a first time sequence feature set; extracting features of structure state data acquired by a sensor to obtain a second time sequence feature set; fusing the first time sequence feature set, the second time sequence feature set and the working condition information to generate a fusion feature vector, inputting the fusion feature vector into a pre-trained hanging basket state prediction model to obtain a future change trajectory and a predicted load distribution; inputting the fusion feature vector, the future change trajectory and the predicted load distribution into a risk assessment model to output a comprehensive risk assessment result; comparing the comprehensive risk assessment result with a preset risk threshold interval to obtain an early warning level, generating a corresponding control instruction and sending the control instruction to a related execution mechanism. The application improves the accuracy and timeliness of construction hanging basket risk monitoring.

Description

Technical Field

[0001] This application relates to the field of monitoring and early warning technology, and in particular to a method and system for real-time monitoring of operational risks of construction hanging baskets. Background Technology

[0002] In the cantilever construction of long-span bridges, the hanging basket serves as the core load-bearing and working platform, and its structural stability directly determines construction safety and project quality. Traditional hanging basket monitoring relies on manual inspections and data from single sensors. This monitoring method suffers from problems such as response lag and incomplete data, making it difficult to cope with multi-source risks in complex construction environments.

[0003] With breakthroughs in computer vision technology, it has become possible to capture the working conditions of the hanging basket in real time through cameras, but this makes up for the lack of intuitiveness of structural data; while the maturity of sensor technology provides data support for obtaining core structural parameters such as stress and displacement.

[0004] Therefore, there is an urgent need for a method and system for real-time monitoring of the operational risks of construction hanging baskets. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a method and system for real-time monitoring of operational risks of construction hanging baskets.

[0006] A first aspect of this application provides a method for real-time monitoring of operational risks of construction hanging baskets, including:

[0007] Video monitoring data is acquired through a camera, and the video monitoring data is analyzed based on computer vision technology to extract the working condition information and the first time-series feature set of the construction hanging basket.

[0008] The structural state data of the construction hanging basket is acquired by sensors, and its features are extracted to obtain a second time-series feature set;

[0009] The first time-series feature set, the second time-series feature set, and the operating condition information are fused to generate a fused feature vector.

[0010] The fused feature vector is input into a deep learning-based hanging basket state prediction model to obtain the future change trajectory of the key parameters of the construction hanging basket and the predicted load distribution of the next construction stage.

[0011] The fused feature vector, the future change trajectory, and the predicted load distribution for the next construction stage are input into the risk assessment model, and a comprehensive risk assessment result is output.

[0012] The comprehensive risk assessment results are compared with the preset risk threshold range to determine the current warning level;

[0013] Based on the warning level, a corresponding generation control command is generated and sent to the relevant execution agency.

[0014] A second aspect of this application provides a real-time monitoring system for the operational risks of construction hanging baskets, comprising:

[0015] The working condition visual perception module is used to acquire video monitoring data through a camera, analyze the video monitoring data based on computer vision technology, and extract the working condition information and the first time-series feature set of the construction hanging basket.

[0016] The structural data acquisition module is used to acquire the structural state data of the construction hanging basket through sensors and extract its features to obtain a second time-series feature set;

[0017] The feature data fusion module is used to fuse the first time-series feature set, the second time-series feature set and the operating condition information to generate a fused feature vector.

[0018] The structural data prediction module is used to input the fused feature vector into the deep learning-based hanging basket state prediction model to obtain the future change trajectory of the key parameters of the construction hanging basket and the predicted load distribution of the next construction stage.

[0019] The comprehensive risk assessment module is used to input the fused feature vector, the future change trajectory, and the predicted load distribution of the next construction stage into the risk assessment model, and output the comprehensive risk assessment result.

[0020] The early warning level determination module is used to compare the comprehensive risk assessment results with a preset risk threshold range to determine the current early warning level;

[0021] The control command execution module is used to generate a corresponding generation control command based on the warning level and send it to the relevant execution mechanism.

[0022] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for real-time monitoring of the operation risk of construction hanging baskets.

[0023] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for real-time monitoring of the operational risks of construction hanging baskets.

[0024] The beneficial effects of the real-time monitoring method and system for construction hanging basket operation risks provided in this application are as follows: This application provides comprehensive protection for the safety management of construction hanging baskets through multi-source data fusion and intelligent analysis. The combined acquisition method of cameras and sensors breaks through the limitations of single data and realizes comprehensive perception of the working condition and structural status of the construction hanging basket, effectively avoiding risk misjudgment caused by data blind spots. The application of computer vision and deep learning technologies improves the accuracy of extracting time-series features and predicting parameter trajectories and load distribution, enabling the hanging basket operation trend to be grasped in advance. In addition, the construction of risk assessment and early warning mechanisms can improve the efficiency of early warning level matching and early warning, effectively shortening the risk response time. At the same time, the automated data processing and control command issuance process ensures the stable operation of the construction hanging basket in complex construction environments. This application improves the accuracy and timeliness of construction hanging basket risk monitoring. Attached Figure Description

[0025] Figure 1 A flowchart illustrating a method for real-time monitoring of operational risks of construction hanging baskets provided in an embodiment of this application;

[0026] Figure 2 A structural block diagram of a real-time monitoring system for the operational risks of construction hanging baskets provided in an embodiment of this application;

[0027] Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0028] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0029] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1-3 The following is an explanation using specific examples.

[0030] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a method for real-time monitoring of operational risks of construction hanging baskets according to an embodiment of this application. The method includes:

[0031] S101: Acquire video monitoring data through a camera, analyze the video monitoring data based on computer vision technology, and extract the working condition information and first time-series feature set of the construction hanging basket.

[0032] In this embodiment, cameras deployed at key locations in the construction area are used to continuously collect video monitoring data throughout the entire hanging basket operation process. The camera deployment is based on the structural characteristics of the hanging basket and the field of view requirements of the construction scene. Unobstructed installation locations that can cover the main load-bearing structure, walking system, and work area of ​​the hanging basket are prioritized to capture core visual information such as the displacement, posture changes, and relative movement of components of the hanging basket.

[0033] Secondly, the collected video monitoring data is analyzed and processed in multiple dimensions based on computer vision technology. Specifically, the video monitoring data is first sampled to obtain multiple consecutive video frames. Video frame preprocessing removes noise such as ambient lighting changes and dust interference, and image enhancement and frame synchronization alignment are applied. Then, the preprocessed video frames are identified to obtain the key components of the construction formwork and their pixel-level contour information. Based on the contour information and camera calibration parameters, the spatial position and attitude angle of the key components are calculated through spatial geometric transformation. Subsequently, a multi-target tracking algorithm is used to temporally correlate the spatial position and attitude angle to obtain the motion trajectory of the key components. The motion trajectory is then analyzed to obtain the motion state and deformation trend of the construction formwork. The motion state, deformation trend, and spatial position and attitude angle are used as the first temporal feature set. The motion state and deformation trend are matched with a preset working condition discrimination rule base to obtain the working condition information of the construction formwork, including static pouring conditions and dynamic movement conditions.

[0034] S102: Obtain structural state data of the construction hanging basket through sensors, extract its features, and obtain the second time-series feature set.

[0035] In this embodiment, based on the mechanical characteristics and risk control priorities of the construction formwork, multiple types of sensors are scientifically deployed at its key locations. Specifically, strain gauges are installed in stress concentration areas of load-bearing components such as the main truss, bottom formwork, and slings of the formwork; displacement sensors are deployed at the formwork's travel track and main node connections; tilt sensors are configured to monitor the overall attitude of the formwork; acceleration sensors capture impact and vibration data; and temperature and humidity sensors can be added in some special scenarios to correct for environmental interference with structural data. The sensor deployment is based on mechanical simulation analysis and verification, ensuring that the sensor coverage extends to the critical stress sections and vulnerable parts of the construction formwork, and that the sensor accuracy meets engineering monitoring requirements. For example, the displacement sensor resolution is less than or equal to 0.01 mm, and the sampling frequency is set to 10-100 Hz.

[0036] The structural state data is preprocessed. For example, filtering algorithms are used to remove redundant information such as vibration interference and electromagnetic noise, and outlier detection and repair algorithms are used to handle data loss or abrupt changes. Subsequently, based on time-series data processing methods, feature extraction is performed on the preprocessed stress, displacement, attitude angle, vibration acceleration, and other data. Specifically, on the one hand, statistical features such as peak value, mean, variance, and peak factor of the structural state data are extracted to represent the static characteristics of the structural state. On the other hand, frequency domain features of the data are mined through methods such as Fourier transform and wavelet analysis to capture the dynamic laws such as the periodicity and attenuation characteristics of structural vibration. Finally, the above multi-dimensional features are summarized to obtain the second time-series feature set.

[0037] S103: Perform data fusion on the first time series feature set, the second time series feature set, and the operating condition information to generate a fused feature vector.

[0038] In this embodiment, multi-source data fusion technology is used to fuse the first time-series feature set, the second time-series feature set, and the working condition information. Specifically, before data fusion, the first time-series feature set, the second time-series feature set, and the working condition information are standardized and preprocessed. For example, for the time-series feature set, Z-Score standardization is used to eliminate the influence of different dimensional parameters such as stress, displacement, and visual displacement. Through a timestamp alignment algorithm, visual features, structural features, and working condition information are unified to the same time scale. The working condition information is quantized and encoded to transform qualitative descriptions such as pouring operations and unloaded walking into computable numerical features.

[0039] During the data fusion process, visual temporal features and structural temporal features are weighted by an attention mechanism to highlight the feature weights of key features such as stress mutations and posture anomalies. Then, the weighted temporal features and quantified working condition information are dimensionally concatenated, and redundant information is removed by principal component analysis to obtain a fused feature vector that includes visual dynamic features, structural mechanical features, and working scene features.

[0040] The feature weights are set adaptively based on an attention mechanism. Specifically, a feature contribution scoring system is established by analyzing the correlation between key features such as stress mutations and abnormal postures in historical data and target events. Target events include structural damage or operational failures. Secondly, the coefficients of the feature weights are adjusted according to the working conditions. For example, when the working condition is pouring concrete, the feature weights of structural temporal features such as stress changes and displacement fluctuations are increased accordingly; when the working condition is unloaded walking, the feature weights of visual temporal features such as visual posture abnormalities and path deviations are increased accordingly. Finally, normalization is performed to ensure that the sum of all feature weights is 1, forming feature weights that match the current working scenario and the state of the construction formwork. A weighted calculation is then performed based on these feature weights to obtain the fused feature vector.

[0041] S104: Input the fused feature vector into the deep learning-based hanging basket state prediction model to obtain the future change trajectory of the key parameters of the construction hanging basket and the predicted load distribution of the next construction stage.

[0042] In this embodiment, the fused feature vector is first subjected to adaptive preprocessing. For example, based on the temporal characteristics of the construction formwork monitoring data and the input requirements of the formwork state prediction model, a sliding time window method is used to reconstruct the feature vector temporally. Specifically, the window length is set to 5-10 sampling intervals according to the construction process cycle of the formwork, which includes the formwork walking cycle and the concrete pouring cycle. Secondly, the preprocessed fused feature vector is subjected to data standardization verification to eliminate the influence of dimensional differences and numerical fluctuations of features of different dimensions.

[0043] After inputting the preprocessed fused feature vector into the hanging basket state prediction model, the model outputs prediction results through multi-dimensional feature mining and temporal pattern learning. These results include the future change trajectory of key parameters of the construction hanging basket and the predicted load distribution for the next construction stage. The future change trajectory of key parameters includes core structural parameters such as main truss stress, sling displacement, overall attitude angle, and node vibration frequency, presented as time-series curves for the next 3-5 construction steps, with confidence interval markers indicating the evolution trend and prediction reliability of each parameter. The predicted load distribution for the next construction stage is based on working condition information and historical load mapping patterns, outputting the load peak value, distribution range, and timing nodes of key loads for each load-bearing component, including the main truss, bottom formwork, and slings.

[0044] This embodiment of the hanging basket state prediction model adopts an encoder-decoder architecture as its core. The encoder consists of two layers of bidirectional gated recurrent units stacked together, responsible for extracting temporal dependencies from the input fused feature vector; the decoder is a single-layer gated recurrent unit used for multi-step prediction. To address the problem of long sequence dependencies, the hanging basket state prediction model integrates a temporal attention mechanism, enabling the decoder to dynamically focus on relevant parts of the encoder's historical sequence at each prediction step. The output layer of the hanging basket state prediction model has a dual-branch structure: one branch regresses and predicts the future change trajectory of key parameters through a fully connected layer, while the other branch outputs the predicted load distribution of each load-bearing component in the next construction stage.

[0045] Training the hanging basket state prediction model is crucial for achieving its prediction function. Specifically, historical fused feature vectors and corresponding future state data are used as training samples, divided into training, validation, and test sets according to a 7:2:1 ratio. Training employs a composite loss function, a weighted sum of the smoothing L1 loss for trajectory prediction and the mean squared error loss for load prediction, minimized using the Adam optimizer. The initial learning rate is 1e-3, and the weight decay is 1e-4. An early stopping mechanism is enabled during training, and performance is monitored on the validation set. For example, if the validation set loss fails to decrease for 10 consecutive training epochs, training is terminated, and the parameters that best reflect the validation set performance are restored.

[0046] The close connection between this hanging basket condition prediction model and the construction hanging basket monitoring scenario is reflected in the meticulous design of its inputs and outputs. The input to the hanging basket condition prediction model is a fusion feature vector integrating visual, structural, and operational information, enabling it to comprehensively perceive the current state of the hanging basket and the operational context. The output of the hanging basket condition prediction model directly serves engineering safety decisions: the future change trajectory of key parameters represents the evolution trend of structural response, providing a forward-looking basis for early warning; the predicted load distribution represents the stress situation of each load-bearing component, used to assess specific risks of load unevenness. Through training, the hanging basket condition prediction model essentially learns the complex mapping relationship between the current multi-source state and future mechanical behavior.

[0047] S105: Input the integrated feature vector, future change trajectory and predicted load distribution of the next construction stage into the risk assessment model, and output the comprehensive risk assessment result.

[0048] In this embodiment, the fused feature vector, the future change trajectory of key parameters, and the predicted load distribution for the next construction stage are used as input data for the risk assessment model. The input data is structured and integrated. Specifically, the fused feature vector serves as the state basis, the future change trajectory represents the key parameters, and the predicted load distribution for the next stage of stress scenarios. These three types of input data provide risk assessment basis from three dimensions: current state, future trend, and work condition prediction.

[0049] Before inputting data into the model, it is necessary to normalize the future change trajectory of key parameters and the predicted load distribution of the next stage of stress scenario in order to eliminate the interference of different dimensions between different data and make the weight ratio of input data in risk assessment reasonable.

[0050] The risk assessment model comprehensively evaluates the operational risks of the construction hanging basket through multi-module collaboration. Specifically, it uses the analytic hierarchy process (AHP) to construct a risk assessment index system, mapping input data to primary indicators such as structural safety, attitude stability, and load adaptability, as well as secondary indicators such as stress exceeding limits, displacement exceeding limits, and uneven load distribution. The weights of each indicator are determined through historical data verification. Then, the fuzzy comprehensive evaluation method is used to quantify the risk level of each indicator. Finally, the model integrates time-series trend risks and predicted risks based on a Bayesian inference model, and outputs a comprehensive risk assessment result including a comprehensive risk score, key risk point location, and risk development trend.

[0051] In this embodiment, the risk assessment model adopts a multi-branch input and feature fusion architecture. Its core modules include: a current state branch, composed of fully connected layers, used to process the fused feature vector representing the real-time health status of the hanging basket; a future trend branch, composed of one-dimensional convolutional layers and fully connected layers, used to analyze the future change trajectory of key parameters obtained from the hanging basket state prediction model and capture potential risk evolution trends; and a working condition prediction branch, composed of fully connected layers, used to analyze the predicted load distribution of the next construction stage and assess the risks under different stress scenarios. The features of each branch are first concatenated, then the interaction relationship between different information sources is explicitly learned through a feature cross-layer, and finally, a comprehensive risk score between 0 and 1 is output through a deep regressor including a dropout layer.

[0052] The risk assessment model is trained using a historical dataset with real-world risk labels. Specifically, during training, mean squared error is used as the loss function, and the Adam optimizer (with an initial learning rate of 5e-4) is employed for parameter updates. The batch size is set to 64, and the training run consists of 200 epochs with an initial learning rate of 5e-4. Through this training process, the risk assessment model can learn complex risk patterns from the data. Specifically, when the current abnormal state, the deteriorating future trend, and the unfavorable load conditions work together, the overall risk score is non-linearly increased.

[0053] The connection between this risk assessment model and the specific scenario of construction formwork is reflected in the design of its inputs and outputs. The three input branches of the risk assessment model correspond to the three core dimensions of risk assessment: current state, future trends, and predicted working conditions. This allows the risk assessment model to conduct comprehensive and forward-looking risk assessments. The final output comprehensive risk score is directly linked to the warning level in engineering practice; for example, a score exceeding 0.8 triggers the highest level of emergency response.

[0054] S106: Compare the comprehensive risk assessment results with the preset risk threshold range to determine the current warning level.

[0055] In this embodiment, a scientifically sound and reasonable pre-set risk threshold interval system is constructed based on the structural safety standards, construction specifications, and historical risk data of the construction formwork. This risk threshold interval system adopts a multi-level division principle, dividing the risk threshold into four intervals, corresponding to four warning levels: Level 1 warning corresponds to a low risk level, with a threshold interval of [0, 0.3], indicating that the construction formwork is stable and meets construction safety requirements; Level 2 warning corresponds to a general risk, with a threshold interval of (0.3, 0.6], indicating that the construction formwork has minor anomalies and requires increased monitoring frequency; Level 3 warning corresponds to a relatively high risk, with a threshold interval of (0.6, 0.8], indicating that the construction formwork has significant potential risks and affects its stability; Level 4 warning corresponds to an extremely high risk, with a threshold interval of (0.8, 1.0], indicating that the construction formwork faces serious safety risks and requires immediate cessation of construction and emergency measures. The division of the threshold intervals was determined through mechanical simulation analysis, expert consultation, and engineering practice verification.

[0056] After constructing the preset risk threshold range, the comprehensive risk assessment result is compared with the aforementioned threshold range to obtain the comparison result. During the comparison process, a numerical matching algorithm is used to determine the specific threshold range in which the comprehensive risk assessment result falls, and then maps it to the corresponding warning level. For example, if the comprehensive risk assessment score is 0.45, it falls within the (0.3, 0.6] range, corresponding to a level 2 warning; if the score is 0.85, it falls within the (0.8, 1.0] range, corresponding to a level 4 warning. Simultaneously, to avoid misjudgments caused by the comprehensive risk assessment result being close to the threshold boundary, a buffer threshold of ±0.02 is set. When the comprehensive risk assessment result falls within the buffer range, a secondary verification mechanism is automatically triggered. For example, based on the key parameter data and future change trajectory data in the fused feature vector, a second verification is performed. Specifically, the data of key parameters such as the main truss stress mutation amplitude, sling displacement rate, and overall attitude angle offset in the fused feature vector are extracted again and compared with the safety threshold for this construction stage. The process involves several steps: first, verifying whether key structural parameters exceed limits in a single dimension or exhibit multi-dimensional synergistic anomalies; second, analyzing the future change trajectories of key parameters to determine the parameter evolution trends corresponding to the comprehensive risk assessment results within the buffer threshold range. For example, if the parameters of a key component show a convergence trend within the next 3-5 construction steps, it is determined to be a false alarm, maintaining the original warning level but increasing the monitoring frequency for that key component; if the parameters of a key component show a continuous deterioration trend, it is determined to be a real risk, and the warning level is raised by one level; finally, based on historical risk data of the construction hanging basket working conditions, and by using a similarity matching algorithm to find historical cases under similar working conditions, the rationality of the comprehensive risk assessment results is assessed, leading to a secondary review conclusion.

[0057] Finally, based on the comparison results and the conclusions of the second review, the warning level of the construction hanging basket was determined and a warning level was generated simultaneously. The warning level includes the risk level name, the corresponding threshold range, the core risk points, and preliminary handling suggestions.

[0058] S107: Generate a corresponding generation control command based on the warning level and send it to the relevant execution agency.

[0059] In this embodiment, a tiered response control command generation mechanism is constructed based on the determined warning level. The control commands are set around a three-dimensional logic of risk level, response intensity, and execution target, with differentiated control commands for different warning levels: Level 1 warning corresponds to Level 1 control commands, which include maintaining the existing monitoring frequency and construction rhythm, and synchronously recording the status data of the construction formwork; Level 2 warning triggers Level 2 control commands for enhanced monitoring and parameter fine-tuning, which include requiring the sensor sampling frequency to be increased to 1.5-2 times the original frequency, and simultaneously optimizing and adjusting construction parameters such as the formwork's travel speed and load application rhythm based on the future change trajectory of key parameters; Level 3 warning generates Level 3 control commands for construction deceleration and special inspections, which mandate the suspension of high-risk work procedures and arrange for technical personnel to conduct on-site verification of key load-bearing components and connection nodes; Level 4 warning corresponds to Level 4 control commands for emergency shutdown and emergency response, which immediately cut off the formwork's power system, activate safety protection devices, organize the evacuation of workers to a safe area, and simultaneously trigger emergency rescue procedures.

[0060] In this embodiment, the secondary control command optimizes and adjusts construction parameters such as the traveling speed of the formwork and the load application rhythm based on the future change trajectory of key parameters. Specifically, it analyzes the future change trajectory of key parameters and extracts core features such as the timing node of the main truss stress peak, the cumulative rate of sling displacement, and the overall attitude angle stability range. When the future change trajectory of key parameters shows that the stress within the next 3-5 construction steps is slowly increasing but still within a safe range, or the formwork displacement rate is within a controllable growth range, the construction parameters are adjusted accordingly. Specifically, for the traveling speed of the formwork, when the predicted trajectory indicates that the attitude angle deviation will approach the critical value, the speed is adjusted according to 0.2-0.3 m / min. The walking speed is gradually reduced while the walking interval is extended, for example, pausing for 30-60 seconds after every 5-8 meters of walking, until the posture angle returns to the stable range before resuming the original speed. For the rhythm of load application, such as concrete pouring, when the predicted stress peak will reach a high level in the middle of the pouring, a segmented pouring + gradient loading mode is adopted, dividing the single pouring volume into 3-4 batches, with the interval between each batch set to 1-2 hours, and the pouring rate of the next batch is dynamically adjusted according to the stress dissipation trajectory of the previous batch. All adjustment actions are subject to hard constraints based on the safety threshold of key parameters, ensuring that after the construction parameters are optimized, the future change trajectory of key parameters is always within the secondary warning threshold range of [0, 0.6].

[0061] Depending on the type of executing agency (such as a hanging basket control system, construction dispatch center, safety monitoring platform, field terminal equipment, etc.), a suitable communication protocol is used to send control commands in a standardized format to the corresponding terminal. The control command content includes key information such as warning level, execution requirements, and time nodes. Simultaneously, a real-time feedback verification process is initiated after the control command is sent, requiring each executing agency to return a control command reception confirmation signal. If no feedback is received within a preset time, the backup communication channel is automatically activated for retransmission, and a manual reminder mechanism is triggered.

[0062] The preset time is based on the real-time requirements of the construction hanging basket control command transmission, the response characteristics of the actuator, and the redundancy setting of the communication environment at the construction site.

[0063] As can be seen from the above, this application provides comprehensive protection for the safety management of construction formwork through multi-source data fusion and intelligent analysis. The combined acquisition method of cameras and sensors breaks through the limitations of single data sources, and simultaneously achieves comprehensive perception of the working conditions and structural status of the construction formwork, effectively avoiding risk misjudgment caused by data blind spots. Furthermore, the application of computer vision and deep learning technologies improves the accuracy of extracting temporal features and predicting parameter trajectories and load distribution, allowing the operating trend of the formwork to be grasped in advance. In addition, the construction of risk assessment and early warning mechanisms can improve the efficiency of early warning level matching and early warning, effectively shortening the risk response time. Simultaneously, the automated data processing and control command issuance process ensures the stable operation of the construction formwork in complex construction environments. This application improves the accuracy and timeliness of risk monitoring for construction formwork.

[0064] In one embodiment of this application, video monitoring data is analyzed based on computer vision technology, and the working condition information of the construction hanging basket and a first temporal feature set are extracted, including:

[0065] Frame sampling processing is performed on the video monitoring data to obtain multiple consecutive video frames;

[0066] The video frames are identified to obtain the key components of the construction hanging basket and their pixel-level contour information is acquired.

[0067] Based on contour information and camera calibration parameters, the spatial position and attitude angle of key components are calculated through spatial geometric transformation.

[0068] Based on a multi-target tracking algorithm, spatial position and attitude angle are correlated in time to obtain the motion trajectory of key components;

[0069] By analyzing the motion trajectory, the motion state and deformation trend of the construction formwork can be obtained;

[0070] Motion state, deformation trend, spatial position and attitude angle are used as the first temporal feature set;

[0071] The motion state and deformation trend are matched with the preset working condition discrimination rule library to obtain the working condition information of the construction formwork, which includes static pouring working condition and dynamic moving working condition.

[0072] In this embodiment, firstly, the video monitoring data collected by the camera is processed by frame sampling. Specifically, based on the dynamic characteristics of the construction hanging basket operation, a frame sampling strategy is adopted for processing. For example, sampling is performed at the original frame rate of 30fps under dynamic movement conditions; under static pouring conditions, the sampling interval is automatically adjusted to 5-10fps, resulting in multiple continuous and temporally coherent video frames. Secondly, the video frames are preprocessed and target recognition is performed. For example, Gaussian filtering and adaptive threshold segmentation algorithms are used to remove interference such as changes in ambient light, construction dust, and background debris. Subsequently, image enhancement technology is used to improve the contrast between the hanging basket components and the background, and target detection algorithms are used to identify key components such as the main truss, bottom formwork, hoisting belts, and traveling wheel sets. At the same time, the Mask R-CNN algorithm is used to generate pixel-level contour information of each component.

[0073] After obtaining pixel-level contour information, the conversion from pixel coordinates to world coordinates is completed through a spatial geometric transformation algorithm based on the camera's calibration parameters. The camera's calibration parameters include intrinsic matrix, distortion coefficients, extrinsic matrix, etc. For example, the three-dimensional spatial coordinates of the contour key points are calculated based on the pixel coordinates and the camera imaging model. Then, the spatial position and attitude angle of each key component are calculated through vector operations, rotation matrix solving, and other methods. The spatial position is represented by X, Y, and Z coordinates in the construction coordinate system. The attitude angles include pitch angle, roll angle, and yaw angle.

[0074] This embodiment uses a multi-target tracking algorithm to temporally correlate spatial position and attitude angle to obtain the motion trajectory of key components. This trajectory is then analyzed. For example, by calculating parameters such as displacement, velocity, acceleration, and curvature of the trajectory, the overall motion state of the construction formwork is determined. Simultaneously, by comparing trajectory deviations of different time periods and different components, and based on the mechanical properties of the components, the presence of abnormal deformation in the construction formwork is analyzed, revealing its deformation trend. Abnormal deformation includes bending of the main truss or tilting of the slings. The motion state and deformation trend of the construction formwork, along with the spatial position and attitude angle of each key component, are integrated to obtain the first temporal feature set.

[0075] Simultaneously, the motion state and deformation trend data of the construction formwork are input into a preset working condition discrimination rule library for matching to obtain the working condition information of the construction formwork. The working condition discrimination rule library is built based on construction process standards and historical working condition data, including discrimination conditions such as static pouring working conditions and dynamic movement working conditions, providing scenario-based basis for subsequent risk assessment.

[0076] As can be seen from the above, this embodiment uses frame sampling processing and computer vision algorithms, which not only ensures a balance between capturing motion details under dynamic working conditions and computational efficiency under static working conditions, but also transforms visual data into quantified three-dimensional parameters such as spatial position and attitude angle through pixel-level contour recognition and spatial geometric transformation. Furthermore, multi-target tracking and trajectory analysis comprehensively uncover the motion state and deformation trend of the hanging basket, not only constructing a multi-dimensional and temporal feature system, but also achieving accurate classification of static pouring and dynamic moving working conditions through a working condition discrimination rule base, thereby improving the automation and intelligence level of hanging basket monitoring, while also reducing the subjectivity and lag of manual monitoring.

[0077] In one embodiment of this application, a multi-target tracking algorithm is used to temporally correlate spatial position and attitude angle to obtain the motion trajectory of a key component, including:

[0078] Kalman filtering is used to predict the position and state of key components in the next frame;

[0079] The predicted location is temporally correlated with the location of key components in the current frame using the Hungarian algorithm to obtain the correlation result.

[0080] Based on the correlation results, the motion state of the key components is updated, and the continuous position coordinates of the key components in the time series are generated to obtain the motion trajectory of the key components.

[0081] In this embodiment, the Kalman filter algorithm is a recursive optimization algorithm based on probability statistics and linear algebra. In the dynamic positioning scenario of key components, it uses multi-frame continuously acquired time-series data as input to construct a state-space model that comprehensively represents the motion characteristics of the key components. Its input data includes motion parameters such as the three-dimensional spatial coordinates, instantaneous velocity, and rate of change of acceleration of the key components in historical frames. These data need to undergo timestamp alignment and noise preprocessing before input. In the prediction phase, the Kalman filter algorithm is based on a preset linear system state equation and the physical laws governing the motion of the key components, such as the law of inertia and kinematic constraints. It also uses matrix operations to deduce the initial position state of the components in the next frame. The state equation describes the evolution of the component's motion state through a state transition matrix, and outputs preliminary results including predicted position, predicted velocity, and the corresponding error covariance matrix.

[0082] Next, in the correction phase, the Kalman filter algorithm utilizes the observed position data of key components in the next frame and calibrates the prediction results through the observation equation. The observation equation establishes a mapping relationship between the physical location and sensor observations using an observation matrix, while also considering observation noise (represented by the observation covariance matrix, quantifying sensor measurement errors). By calculating the Kalman gain, the predicted and observed positions are weighted and fused to correct the initial prediction error, ultimately resulting in the predicted position of the key component in the next frame.

[0083] Subsequently, the predicted location is temporally correlated with the location of the key component in the current frame using the Hungarian algorithm to obtain the correlation result. Based on the correlation result, the motion state parameters of the key component are corrected and iteratively updated. At the same time, based on the continuity constraint of the time series, the spatial position coordinates of the key component at each time node are generated through temporal interpolation and state prediction algorithms. Finally, the motion trajectory of the key component is formed by connecting the two together, covering the entire observation period, with continuous coordinates and consistent motion patterns.

[0084] The input to the multi-target tracking algorithm is the spatial position and attitude angle of each key component, which is extracted from continuous video frames using computer vision technology and obtained based on spatial geometric transformation.

[0085] The core processing logic of the multi-target tracking algorithm aims to solve problems such as target occlusion, appearance changes, and complex background interference in construction scenarios. It employs a Kalman filter algorithm to predict the position and state of components in the next frame based on historical motion data and constructs their motion trajectories. Then, using the Hungarian algorithm, it constructs a cost matrix by calculating the intersection-union ratio, center point distance, and shape similarity between the predicted position and the actual detected position in the current frame, and performs optimal matching to achieve temporal correlation of the same component across different frames. The output of this multi-target tracking algorithm is the continuous motion trajectory of each key component over time. This trajectory data not only includes the historical sequence of the key component's position and attitude, but more importantly, by analyzing the trajectory, the overall motion state and deformation trend of the construction formwork can be further obtained.

[0086] Specifically, in the visual monitoring scenario of a construction hanging basket, the Kalman filter is constructed to establish a kinematic model for each tracked key component. Its input is the state vector of the key component in historical frames, typically including its position, velocity, and even acceleration information in three-dimensional space. For example, the state vector can be set as x=[p x ,p y ,p z ,v x ,v y ,v z ] THere, p represents position and v represents velocity. The core layers of the Kalman filter include a prediction step and an update step. For example, the prediction step predicts the prior state estimate of the component in the next frame based on the linear state transition equation and the error covariance matrix; the update step uses the actual observations of the current frame and corrects the prior estimate based on the Kalman gain to obtain the optimal posterior state estimate. Key parameters include the process noise covariance matrix Q and the observation noise covariance matrix R. Q is set according to the possible perturbations of the basket's motion, while R is determined based on camera calibration and image recognition accuracy. The output of the Kalman filter is a smoother and more accurate predicted position and a measure of its uncertainty.

[0087] The Hungarian algorithm aims to solve the problem of optimally allocating detected targets to existing tracking trajectories between frames. Its input is a cost matrix, which is composed of pairwise matching costs calculated from multiple target positions predicted by Kalman filtering and multiple newly detected target positions in the current frame. The matching cost integrates multiple metrics such as intersection-union ratio (IU), Euclidean distance between center points, and shape similarity, assigning them different weights. These costs are then normalized to comparable dimensions, and a preset threshold is used to filter out unreasonable matching pairs, marking them as invalid. As a combinatorial optimization algorithm, the core of the Hungarian algorithm is to traverse the cost matrix, using row and column reduction steps to find an optimal set of matching pairs that minimizes the overall matching cost. Its output is a list of optimal matching pairs, along with unmatched critical components and newly appearing unmatched critical components.

[0088] As can be seen from the above, the Kalman filter in this embodiment can accurately predict the position status of key components in the next frame based on historical motion data, effectively suppressing the interference of environmental noise and measurement errors on position detection; at the same time, it generates motion trajectories based on the update of motion status based on the correlation results, thereby improving the stability of visual monitoring of construction hanging baskets.

[0089] In one embodiment of this application, the predicted location is temporally correlated with the location of a key component in the current frame using a Hungarian algorithm to obtain the correlation result, including:

[0090] Calculate the intersection-union ratio, center point distance, and shape similarity between the predicted location and the location of key components in the current frame, and construct them as a cost matrix;

[0091] The cost matrix is ​​normalized to obtain the standard cost matrix;

[0092] Matches in the standard cost matrix with an intersection-union ratio less than a preset first threshold are marked as invalid matches;

[0093] The Hungarian algorithm is used to solve for the optimal matching of the processed standard cost matrix to obtain matching pairs.

[0094] Based on the matching pairs and invalid matches, distinguish between successfully associated key components, newly emerging key components, and key components that failed to be matched, and obtain the association results.

[0095] In this embodiment, three core matching metrics are calculated between the predicted location and the location of key components in the current frame to construct a cost matrix. Specifically, the Intersection over Union (IoU) measures the degree of overlap between the predicted and detected regions, representing the consistency of spatial positions; the center point distance is calculated using Euclidean distance to quantify the degree of deviation between the two locations; and shape similarity is determined based on parameters such as the aspect ratio of the bounding rectangle of the pixel-level contour and the contour area ratio to assess the morphological matching degree between the predicted location and the key component location. The three metrics are weighted and fused according to preset metric weights to obtain the cost matrix. The IoU accounts for 50%, the center point distance for 30%, and the shape similarity for 20%.

[0096] Secondly, the Min-Max normalization method is used to map the intersection-union ratio (IU), centroid distance, and shape similarity to the [0,1] interval. The IU and shape similarity are mapped forward, meaning a larger value indicates a higher matching degree. The centroid distance is mapped backward, meaning a smaller value indicates a higher matching degree. The normalized cost matrix is ​​used as the standard cost matrix. Matches in the standard cost matrix with an IU less than a preset first threshold are marked as invalid matches.

[0097] Subsequently, the Hungarian algorithm is used to solve for the minimum weight and the corresponding matching combination by traversing all non-invalid markers in the matrix, and output the matching pair that satisfies the global optimum.

[0098] Based on the matching pairs and invalid matches, the association results are obtained. The association results include key components that have been successfully associated, newly emerging key components, and key components that failed to be matched.

[0099] As can be seen from the above, this embodiment constructs a cost matrix using three core indicators: intersection-union ratio, center point distance, and shape similarity, thus avoiding misjudgments caused by a single indicator. Simultaneously, normalization eliminates dimensional differences, and invalid matching terms are filtered based on a first threshold, effectively reducing the impact of interference factors. Furthermore, the Hungarian algorithm is used to solve for the globally optimal matching pair, distinguishing key components of different categories, improving the system's adaptability to complex scenarios, and ensuring the integrity and accuracy of the association results.

[0100] In one embodiment of this application, before updating the motion state of the key component based on the correlation result and generating the continuous position coordinates of the key component in the time series to obtain the motion trajectory of the key component, the method further includes:

[0101] For a key component that is successfully matched, its state and motion trajectory are updated using its position in the current frame.

[0102] For critical components that fail to match, if they are detected in multiple consecutive frames and the confidence level is greater than the preset confidence threshold, they are identified as newly emerging critical components and their motion trajectories are created.

[0103] If a critical component fails to match successfully within a preset first time window, it is determined that the critical component has left the monitoring area, and tracking of it is terminated.

[0104] In this embodiment, for a key component that is successfully matched, the motion state parameters of the component are updated using its detection position in the current frame, and the position coordinates of the current frame are appended to the historical trajectory data of the key component in order of timestamp.

[0105] For critical components that fail to match, a dual verification mechanism is initiated to determine their status. For example, firstly, the detection record of the critical component is checked. If it is consistently detected within a consecutive number of frames, and the detection confidence level for each frame is greater than a preset confidence threshold, it is identified as a newly emerging critical component. Subsequently, its motion state parameters are initialized, and a tracking link and trajectory data storage structure is created. The coordinates of the newly emerging critical component's position in the current and historical detection frames are entered to obtain its motion trajectory. The consecutive number of frames is generally 3-5 frames, adjusted according to the construction hanging basket working scenario. The preset confidence threshold is determined based on engineering practice verification. This preset confidence threshold is determined based on the distribution characteristics of critical component detection data in the construction hanging basket monitoring scenario, the intensity of environmental interference, and the balance requirement between false positive and false negative rates in historical detection cases in engineering practice.

[0106] Key components that fail to match include situations where the original tracking component fails to find a match or the newly detected component does not meet the criteria for new appearance. For key components that consistently fail to match, a timeout determination is made based on a preset first time window, which is calculated according to the camera sampling frequency. For example, if a key component that consistently fails to match fails to successfully match any predicted or detected location within the first time window, and no valid detection record is updated, then it is determined that the key component has left the monitoring area or has ceased to be a key monitoring target. In this case, the tracking process for that key component is terminated.

[0107] As can be seen from the above, this embodiment updates the status and trajectory of successfully matched components, ensuring the continuity and accuracy of time-series data; while components that are not matched but are detected stably for multiple consecutive frames and meet the confidence level are identified as new targets and have their trajectories created, avoiding the problem of missing key monitoring objects; at the same time, tracking of components that are continuously unmatched within the first time window is terminated, reducing the occupation of ineffective computing resources.

[0108] In one embodiment of this application, after marking matches in the standard cost matrix whose intersection-union ratio is less than a preset first threshold as invalid matches, the method further includes:

[0109] Acquire auxiliary sensor data synchronized with video monitoring data, including laser ranging point cloud or millimeter-wave radar data;

[0110] Spatial location verification is performed between the predicted location of the key component corresponding to the invalid match and the auxiliary sensor data.

[0111] If the spatial location verification passes, the corresponding invalid match is corrected to a valid match and re-included in the standard cost matrix;

[0112] If the spatial location verification fails, it will remain an invalid match and its corresponding key component will be marked as a suspected interference source.

[0113] In this embodiment, auxiliary sensor data synchronized with video monitoring data is first acquired. Specifically, based on the monitoring needs and environmental adaptability of the construction scenario, laser ranging point cloud or millimeter-wave radar data is selected as a supplementary verification data source, and the auxiliary sensor data and video monitoring data are time-aligned using a timestamp alignment algorithm.

[0114] For invalid matches already marked in the standard cost matrix, a spatial location verification process is initiated. For example, the predicted location of the key component corresponding to the invalid match is extracted, and the point cloud or radar detection results of the corresponding spatial region are parsed from the auxiliary sensor data. The predicted location and the auxiliary sensor detection location are calculated to obtain the spatial Euclidean distance, and the spatial Euclidean distance is compared with the preset verification threshold.

[0115] For example, if the spatial Euclidean distance is less than or equal to the verification threshold, it means that the predicted position of the invalid match matches the detection result of the auxiliary sensor, and the spatial position verification is deemed successful. At this time, the invalid match is corrected to a valid match and re-included in the standard cost matrix as a cost parameter in the update matrix; if the spatial Euclidean distance is greater than the verification threshold, the spatial position verification is deemed unsuccessful, and the invalid match is maintained as an invalid match; at the same time, its corresponding key component is marked as a suspected interference source.

[0116] From the above, it can be concluded that this embodiment leverages the spatial detection advantages of laser ranging point cloud or millimeter-wave radar data, which complement video monitoring data, to perform secondary verification on invalid matching items. This effectively corrects misjudgments caused by occlusion and light and shadow interference from single visual data. At the same time, by marking suspected interference sources by verifying failed items, the impact of irrelevant factors such as environmental debris on the tracking link is reduced, ensuring the effectiveness of the standard cost matrix.

[0117] In one embodiment of this application, if the spatial location verification fails, it remains an invalid match, and the critical component is marked as a suspected interference source, including:

[0118] Within the preset second time window, suspected interference sources are re-identified and tracked;

[0119] If, within the second time window, the number of times a suspected interference source is successfully matched in consecutive video frames exceeds a preset second threshold, then its suspected interference source label is removed, and it is used as a newly emerging key component to create a new motion trajectory.

[0120] If the number of successful matches of a suspected interference source is still less than the second threshold when the second time window ends, it will be determined to be an interference source.

[0121] In this embodiment, suspected interference sources are re-identified and tracked within a preset second time window. For example, during continuous tracking within the second time window, the number of successful matches of the suspected interference source is counted. If, before the end of the second time window, the number of successful matches of the suspected interference source exceeds a preset second threshold, it is determined that the suspected interference source is not environmental interference, but rather a critical component that has not been identified in time. The suspected interference source is then removed from the flag and processed according to the processing flow for newly emerging critical components. The second threshold is set to 1 to 3 seconds to allow sufficient time to observe the target's recurrence behavior without accumulating excessive noise due to prolonged waiting. The second time window is determined through comprehensive analysis of the system's maximum tolerable delay, target motion characteristics, and historical interference statistics.

[0122] If, after the second time window ends, the number of successful matches for a suspected interference source is still less than or equal to the second threshold, it indicates that the suspected interference source does not possess the stable motion patterns and characteristics of key components, and the suspected interference source is then identified as an interference source. Furthermore, an interference source shielding mechanism is implemented, reducing the detection priority of this interference source in subsequent monitoring processes and removing it from the key component matching range to avoid consuming computing resources or interfering with the tracking of effective targets.

[0123] From the above, it can be concluded that this embodiment provides sufficient verification period for the determination of suspected interference sources based on the second time window, and dynamically monitors suspected interference sources based on multi-dimensional matching standards, avoiding the omission of key components due to single misjudgment; and by quantitatively comparing the number of successful matches with the second threshold, it reduces the occupation of invalid computing resources and improves the accuracy of monitoring key components.

[0124] In one embodiment of this application, the fused feature vector is input into a deep learning-based hanging basket state prediction model to obtain the future change trajectory of key parameters of the construction hanging basket and the predicted load distribution for the next construction stage, including:

[0125] The fused feature vector is input into the encoder of the hanging basket state prediction model, and its temporal features are extracted through a multi-layer gated recurrent unit to obtain the encoded state sequence.

[0126] The encoded state sequence is input into the decoder of the hanging basket state prediction model to obtain the predicted values ​​for multiple future time steps; the predicted values ​​include the instantaneous predicted values ​​of the key parameters and the instantaneous predicted values ​​of the load distribution.

[0127] The instantaneous predicted values ​​of key parameters are sequentially integrated to generate the future change trajectory of key parameters of the construction hanging basket;

[0128] The instantaneous predicted values ​​of the load distribution are sequentially integrated to generate the predicted load distribution for the next construction stage.

[0129] In this embodiment, the fused feature vector is input into the encoder of the hanging basket state prediction model. The encoder uses a multi-layer gated recurrent unit as the core network structure. Utilizing the reset and update gate mechanisms unique to the multi-layer gated recurrent unit, it captures the long-term and short-term temporal dependencies in the fused features. For example, the reset gate can selectively forget redundant information in historical monitoring data, while the update gate adaptively retains temporal features that have a key impact on changes in the hanging basket state. Through deep iterative computation of the multi-layer network, an encoded state sequence that can characterize the evolution law of the hanging basket's operating state is extracted.

[0130] Simultaneously, the encoded state sequence is input into the decoder module of the hanging basket state prediction model. Based on the temporal feature representation output by the encoder, the decoder focuses on key temporal nodes using an attention mechanism and outputs instantaneous prediction results for multiple consecutive future time steps through a multi-step prediction strategy. These prediction results specifically include two types of core data: first, instantaneous predicted values ​​of the hanging basket's key parameters, representing the specific values ​​of each parameter at a future specific moment; and second, instantaneous predicted values ​​of the hanging basket's load distribution, characterizing the load allocation across different load-bearing parts of the hanging basket at different times.

[0131] A time-series integration algorithm is used to process the instantaneous predicted values ​​of key parameters. Specifically, by performing trend fitting, outlier correction, and smoothing on the instantaneous prediction results of multiple future time steps, random errors in single-step predictions are eliminated, resulting in the future change trajectory of the key parameters. This future change trajectory can not only represent the changing trend, peak range, and stable range of each parameter, but also help construction personnel to predict in advance whether the parameters will be abnormal.

[0132] This embodiment integrates load distribution data from multiple time steps to generate a predicted load distribution for the next construction stage, which is presented as a prediction map. This prediction map is used to guide the optimization of the construction plan, such as adjusting the layout of the hanging basket support points and optimizing the load distribution strategy.

[0133] As can be seen from the above, this embodiment deeply extracts the temporal features of the fused feature vector through a multi-layer gated loop unit, and predicts the key parameters and load distribution for multiple future time steps based on the encoder-decoder architecture, generating the key parameter change trajectory and load distribution for the next construction stage. This not only effectively captures the temporal dependency and dynamic change law of the construction hanging basket state evolution, but also improves the accuracy and timeliness of the prediction results.

[0134] In one embodiment of this application, the encoded state sequence is input to the decoder of the hanging basket state prediction model to obtain predicted values ​​for multiple future time steps, including:

[0135] At each prediction time step of the decoder, the association weight between the current decoding time step and the historical encoding time steps is calculated through a temporal attention mechanism to obtain the association weight;

[0136] The encoder's hidden states are weighted and summed based on the association weights to obtain the encoder state sequence at the current time step.

[0137] The encoded state sequence of the current time step is concatenated with the output of the decoder from the previous time step.

[0138] The concatenated features are nonlinearly transformed by a fully connected layer to output the predicted value at the current time step.

[0139] When the prediction for the preset future time step is reached, the predicted values ​​for multiple future time steps are obtained.

[0140] In this embodiment, during the decoder operation of the hanging basket state prediction model, each prediction time step initiates the association weight calculation process with a temporal attention mechanism as its core. For example, the temporal attention mechanism first captures the latent state features of the current decoding time step, and then calculates these latent state features one by one with the latent states of each historical encoding time step output by the encoder, obtaining three types of calculation results. Specifically, the cosine similarity of the feature vectors is obtained by the ratio of the vector dot product to the product of the modulus; the temporal distance decay coefficient is obtained based on the exponential decay function; the contribution of historical states is obtained through correlation analysis between historical states and the current prediction target; subsequently, a weighted summation is performed on the three types of calculation results corresponding to each historical encoding time step based on a preset weight ratio to obtain the initial association score between the historical encoding time step and the current decoding time step; finally, the initial association score is normalized using the Softmax function, mapping the score to the 0-1 interval, and the sum of all scores is 1, ultimately obtaining the association weights corresponding one-to-one between the current decoding time step and each historical encoding time step. The preset weight ratios are determined as follows: based on the time-series prediction characteristics of key parameters in the hanging basket construction scenario, and obtained through a combination of offline experiments and engineering data calibration. For example, a validation set is first constructed based on historical monitoring data, and different weight combinations are traversed using grid search or cross-validation methods. The optimal weight combination is selected with the goal of minimizing prediction error. At the same time, the weight allocation is referenced from the experience of similar time-series prediction tasks in the engineering field, and manual correction is performed according to the actual impact priority of feature matching degree, time-series correlation, and historical validity in the hanging basket status data to obtain the weight ratios.

[0141] The encoder outputs a weighted sum of all historical hidden states based on association weights to obtain the encoded state sequence at the current time step. The association weight of the i-th historical encoded time step corresponds to the i-th historical hidden state output by the encoder. Next, the encoded state sequence at the current time step is concatenated with the output of the decoder from the previous time step. Specifically, after generating the encoded state sequence at the current time step, the feature dimension concatenation method is used to fuse it with the output of the decoder at the previous time step: First, the dimensionality consistency of the two types of features is checked. If the dimension of the encoded state sequence is d and the dimension of the output at the previous time step is m, and d≠m, then the output of the previous time step is mapped to the d dimension through a 1×1 convolutional layer, or the dimension of the encoded state sequence is reduced to the m dimension through a fully connected layer, so that the dimensional format of the two types of features is unified. Then, the two types of features are integrated through vector concatenation to obtain a composite feature vector. The encoded state sequence retains the key temporal patterns and effective information in the historical hidden states, while the output of the decoder at the previous time step carries the real-time dynamic feedback of the prediction process. The concatenated composite feature vector not only fully retains the original information features of the two types of inputs, but also realizes the organic integration of historical encoded features and real-time prediction feedback.

[0142] In this embodiment, the concatenated composite feature vector is input into a fully connected layer, and a nonlinear transformation is performed through a multilayer perceptron structure to obtain the predicted value for the current time step. Specifically, the activation function is deployed in the hidden layer of the fully connected layer to mine complex nonlinear correlations between features and capture deep feature interaction patterns. Then, using the dimension mapping mechanism of the output layer, the nonlinearly transformed feature vector is projected onto a dimension space that matches the prediction target of the hanging basket state, finally obtaining the predicted value for the current time step. The prediction target of the hanging basket state includes key parameters such as the main truss stress and the displacement of the slings.

[0143] The above process is executed cyclically at each prediction time step until the preset threshold for the number of future time steps is reached. For example, if the preset future prediction time steps are 3-5 construction steps, the hanging basket state prediction model will sequentially complete the association weight calculation, encoding state fusion, feature splicing and nonlinear transformation for each time step, and finally output continuous prediction values ​​for multiple future time steps, forming a complete time series prediction sequence.

[0144] As can be seen from the above, this embodiment calculates the association weights through a temporal attention mechanism, enabling the hanging basket state prediction model to focus on historical feature information with higher prediction value and effectively filter redundant data. The weighted summation operation based on the association weights achieves differentiated fusion of historical hidden states, and the generated encoded state sequence is more targeted. At the same time, the nonlinear transformation of the fully connected layer can deeply mine feature associations, and finally obtain continuous future results through multi-time step cyclic prediction. This series of operations improves the accuracy, temporal coherence, and engineering practicality of the prediction, providing a reliable temporal prediction basis for scenarios such as hanging basket state monitoring.

[0145] In one embodiment of this application, a method for real-time monitoring of operational risks of construction hanging baskets further includes:

[0146] The error between the predicted value and the monitored value is calculated to obtain the prediction error of the hanging basket state prediction model;

[0147] The temperature coefficient in the temporal attention mechanism and the forgetting gate bias parameter in the gated loop unit are adjusted based on the changing trend of the prediction error; whereby...

[0148] Adjusting the temperature coefficient based on the trend of prediction error includes:

[0149] When the trend is upward, use the preset first step length to reduce the temperature coefficient;

[0150] When the trend is downward or stable, the temperature coefficient is increased using the preset second step size.

[0151] Adjusting the forget gate bias parameters based on the trend of prediction error changes includes:

[0152] When the prediction target is the long-term future trajectory of key parameters, the forget gate bias parameter is increased using a preset third step size.

[0153] When the prediction target is the instantaneous predicted value of the load distribution, the forget gate bias parameter is reduced using the preset fourth step size.

[0154] In this embodiment, the error between the predicted value output by the hanging basket state prediction model and the monitored value is first calculated, such as mean square error or mean absolute error, to obtain the prediction error of the hanging basket state prediction model. The predicted values ​​include the instantaneous predicted values ​​of key parameters, the instantaneous predicted values ​​of load distribution, and the predicted results of the long-term future change trajectory of key parameters. Based on the chronological order of the prediction time steps, the prediction errors corresponding to each time step are recorded sequentially to form a continuous error data sequence. The error data sequence is then analyzed to obtain the trend of the prediction error.

[0155] Based on the changing trend of the prediction error, the hanging basket state prediction model specifically adjusts the temperature coefficient in the temporal attention mechanism and the forget gate bias parameter in the gating loop unit.

[0156] Specifically, the adjustment of the temperature coefficient includes: when the prediction error shows an upward trend, it indicates that the current hanging basket state prediction model's screening accuracy for historical features is insufficient. In this case, the temperature coefficient is reduced by a preset first step length, making the weight distribution more concentrated on core historical features, thereby reducing redundant information interference. When the prediction error shows a downward or stable trend, it indicates that the current feature screening logic of the hanging basket state prediction model is effective. In this case, the temperature coefficient is increased by a preset second step length, appropriately expanding the scope of attention for historical features and enhancing the hanging basket state prediction model's ability to capture the state of the construction hanging basket.

[0157] The adjustment of the forget gate bias parameter in the gated loop unit includes: when the prediction target is the long-term future change trajectory of key parameters, it is necessary to enhance the basket state prediction model's ability to remember long-term historical time series patterns. Therefore, based on the prediction error change trend, the forget gate bias parameter is increased using a preset third step size, and the opening degree of the forget gate is reduced, so that the basket state prediction model retains more historical time series characteristics. When the prediction target is the instantaneous prediction value of load distribution, it is necessary to improve the response speed of the basket state prediction model to changes in operating conditions. Therefore, also combined with the prediction error change trend, the forget gate bias parameter is decreased using a preset fourth step size, and the opening degree of the forget gate is increased, so that the basket state prediction model's learning and adaptation to the latest operating condition characteristics are enhanced.

[0158] The first step length is based on the need to quickly suppress the error when the prediction error rises, and is a value preset based on historical error adjustment data and engineering experience to reduce the temperature coefficient. Its absolute value is usually greater than the second step length to ensure adjustment sensitivity.

[0159] The second step is based on the goal of steadily optimizing model performance when the prediction error decreases or stabilizes. It is a preset value for improving the temperature coefficient by referring to the experience of adjusting parameters of similar time series prediction models, so as to achieve a gradual expansion of the feature focus range.

[0160] The third step is to enhance the memory of historical information in order to predict the long-term change trajectory of key parameters. It is to increase the value of the forget gate bias parameter by combining the forget gate adjustment law of the gated cycle unit with the requirements of long-term prediction tasks.

[0161] The fourth step is to improve the real-time response capability in the instantaneous prediction of load distribution. Based on the processing characteristics of the forget gate for instantaneous information and the engineering measured data, the preset value of the forget gate bias parameter is calibrated to reduce the forget gate bias parameter.

[0162] As can be seen from the above, this embodiment improves the accuracy and adaptability of the hanging basket state prediction model by adaptively optimizing the core parameters driven by the prediction error. Adjusting the temperature coefficient based on the error change trend and differentiating the forgetting gate bias parameter according to the prediction target effectively improves the stability and accuracy of the hanging basket state prediction model, providing reliable technical support for the safety monitoring of hanging basket construction.

[0163] In one embodiment of this application, a method for real-time monitoring of operational risks of construction hanging baskets further includes:

[0164] The monitoring method was tested under various extreme environmental conditions to build a validation dataset. Extreme environmental conditions included heavy fog, rainstorms, dust storms, and physical obstructions that could degrade the quality of video monitoring data.

[0165] The monitoring data acquired under various extreme environmental conditions were analyzed using computer vision technology, and features were extracted to obtain a third time-series feature set.

[0166] Based on the validation dataset, the success rate of extracting the third temporal feature set was evaluated, and the performance degradation curve of computer vision technology under different environmental conditions was quantified.

[0167] Based on the performance degradation curve, the fusion weights of the first time series feature set and the second time series feature set are adjusted.

[0168] In this embodiment, the monitoring method is first run under various extreme environmental conditions to construct a validation dataset. Specifically, extreme environmental conditions include typical scenarios that lead to a decline in the quality of video monitoring data, such as visual blurring in foggy weather, lens obstruction and image noise caused by heavy rain, reduced visibility caused by dust at construction sites, and physical obstruction caused by construction equipment or building materials.

[0169] Secondly, monitoring data under various extreme environmental conditions were collected and analyzed and feature extracted using computer vision technology to obtain a third temporal feature set. During data processing, appropriate preprocessing strategies were adopted to address the interference characteristics of different extreme environments, such as defogging algorithms for foggy scenes, noise filtering for rainstorm scenes, and feature completion for occluded scenes. Subsequently, core temporal features such as displacement and deformation of the construction formwork structure were extracted using computer vision technologies such as target detection and key point tracking, ensuring the effectiveness and completeness of the third temporal feature set.

[0170] Subsequently, based on the validation dataset, the success rate of extracting the third temporal feature set was comprehensively evaluated. In practice, the monitoring data in the validation dataset were classified according to environmental types such as heavy fog, heavy rain, dust storms, and physical occlusion. Then, for each type of environment, the samples extracted by computer vision technology were individually verified to ensure completeness and consistency with real-world conditions. The proportion of valid feature samples in each environment to the total number of samples was calculated as the feature extraction success rate for that environment. Simultaneously, the success rate fluctuations under different interference intensities were compared across environments to determine the impact of extreme environments on feature extraction. Furthermore, the performance degradation curves of computer vision technology under different environmental conditions were quantified. These performance degradation curves, plotted with environmental interference intensity on the horizontal axis and feature extraction success rate on the vertical axis, represent the performance degradation patterns of computer vision technology under different harsh environments.

[0171] Finally, based on the performance degradation curve, the fusion weights of the first and second temporal feature sets are dynamically adjusted. Specifically, when extreme environments cause severe performance degradation of computer vision technology, the fusion weights of the second temporal feature set are appropriately increased to reduce dependence on the disturbed visual features; when environmental conditions improve and visual feature extraction performance recovers, the weight ratios are adjusted in the opposite direction.

[0172] As can be seen from the above, this embodiment, by constructing a verification dataset and extracting a third temporal feature set under extreme environments such as heavy fog and torrential rain, can obtain the performance of computer vision technology under harsh working conditions. Furthermore, the quantification performance degradation curve can define the degree of interference of different environments on feature extraction, and then adjust the fusion weights of the first and second temporal feature sets in a targeted manner. When video monitoring data fails, the weight ratio of structural state data is increased, and when the environment improves, the effective contribution of video monitoring data is restored. This not only ensures the integrity and accuracy of monitoring data under extreme environments, but also enhances the robustness and environmental adaptability of the entire monitoring method, providing stable and reliable technical support for real-time early warning of hanging basket operation risks in complex construction scenarios.

[0173] In one embodiment of this application, the fusion weights of the first time-series feature set and the second time-series feature set are adjusted based on the performance degradation curve, including:

[0174] Determine the confidence value of computer vision technology under the current environmental conditions based on the performance degradation curve;

[0175] The weight coefficients of the first time series feature set are calculated based on the confidence scores;

[0176] Based on the weighting coefficients and the preset baseline weights of the sensor data, the complementary weighting coefficients of the second time-series feature set are calculated.

[0177] A weighted fusion algorithm is used to fuse the first time series feature set according to weight coefficients and the second time series feature set according to complementary weight coefficients;

[0178] When the confidence value is less than the preset environmental adaptation threshold, the sensor data-dominated mode is activated, in which the weight of the second time-series feature set is greater than that of the first time-series feature set.

[0179] In this embodiment, firstly, the confidence value of computer vision technology under the current environmental conditions is determined based on the performance degradation curve. Specifically, the performance degradation curve already shows the performance loss pattern of visual feature extraction under different extreme environments and interference intensities. Therefore, the corresponding interval of the curve can be matched according to environmental monitoring data, and the indicators such as feature extraction success rate and accuracy degradation ratio in this interval can be normalized to obtain the confidence value in the 0-1 interval. The smaller the confidence value, the worse the reliability of computer vision technology under the current environment.

[0180] Secondly, the weight coefficients of the first temporal feature set are calculated based on the confidence scores. For example, the first temporal feature set consists of visual temporal features extracted by computer vision technology. Its weight coefficients are positively correlated with the confidence scores. For instance, a linear mapping formula is used to convert the confidence scores into corresponding weight coefficients, ensuring that the stronger the performance of the computer vision technology, the higher the weight of the first temporal feature set in the fusion process; conversely, its weight is reduced to minimize the interference of invalid or erroneous features on the monitoring results. The linear mapping formula is as follows:

[0181] W1 = Wmin + Cv × (Wmax - Wmin)

[0182] Cv is the confidence level value, determined by the performance decay curve, and its range is [0,1]. W1 is the weight coefficient of the first temporal feature set to be calculated. Wmax is the upper limit of the weight of the first temporal feature set, representing the highest weight that visual features can occupy under ideal conditions (Cv=1). Wmin is the lower limit of the weight of the first temporal feature set, representing the lowest weight that visual features must retain under worst-case conditions (Cv=0).

[0183] Subsequently, based on the weighting coefficients and the preset baseline weights of the sensor data, the complementary weighting coefficients of the second time-series feature set are calculated. The baseline weights of the second time-series feature set are set based on engineering experience and the reliability of the sensor data, and are calculated using the formula: 1 - weighting coefficients of the first time-series feature set + correction value of the baseline weights. This process avoids the paralysis of the monitoring system due to the failure of a single feature set.

[0184] Next, a weighted fusion algorithm is used to fuse the first temporal feature set according to weight coefficients and the second temporal feature set according to complementary weight coefficients. During the fusion process, the temporal dimensions of the two feature sets are aligned, and feature complementarity is achieved through algorithms such as weighted averaging, feature concatenation, or attention fusion.

[0185] Finally, when the confidence level is less than the preset environmental adaptation threshold, the sensor data-dominated mode is activated, in which the weight of the second time-series feature set is greater than that of the first time-series feature set. The environmental adaptation threshold is set according to the accuracy requirements of engineering safety monitoring. When the confidence level is less than the environmental adaptation threshold, it indicates that computer vision technology cannot provide effective monitoring data in the current environment. At this time, the weight of the second time-series feature set is increased to a preset dominance ratio, focusing on risk monitoring based on sensor data to ensure the continuity and accuracy of risk monitoring for hanging basket operation in extremely harsh environments. The dominance ratio is above 60%.

[0186] From the above, it can be concluded that this embodiment, by adjusting the weight coefficients of the first and second time-series feature sets based on confidence values, can leverage the advantages of video monitoring data while ensuring the fundamental support role of structural state data through complementary weights. Furthermore, the use of a weighted fusion algorithm achieves organic complementarity between the two types of features, improving the comprehensiveness and reliability of the monitoring data. In addition, the triggering mechanism of the sensor data-dominated mode can switch to stable sensor data dominance when video monitoring data fails, improving the continuity and accuracy of risk monitoring for hanging basket operation in extreme environments.

[0187] Corresponding to the real-time monitoring method for the operational risks of the construction hanging basket in the above embodiment, Figure 2 This is a structural block diagram of a real-time monitoring system for the operational risks of a construction hanging basket, provided in one embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 2 The real-time monitoring system 20 for the operation risk of the construction hanging basket includes: a working condition visual perception module 21, a structural data acquisition module 22, a feature data fusion module 23, a structural data prediction module 24, a comprehensive risk assessment module 25, a warning level determination module 26, and a control command execution module 27.

[0188] The working condition visual perception module 21 is used to acquire video monitoring data through a camera, analyze the video monitoring data based on computer vision technology, and extract the working condition information and the first time-series feature set of the construction hanging basket.

[0189] The structural data acquisition module 22 is used to acquire structural state data of the construction hanging basket through sensors and extract its features to obtain a second time-series feature set;

[0190] The feature data fusion module 23 is used to fuse the first time series feature set, the second time series feature set and the operating condition information to generate a fused feature vector.

[0191] The structural data prediction module 24 is used to input the fused feature vector into the deep learning-based hanging basket state prediction model to obtain the future change trajectory of the key parameters of the construction hanging basket and the predicted load distribution of the next construction stage.

[0192] The comprehensive risk assessment module 25 is used to input the fused feature vector, future change trajectory and predicted load distribution of the next construction stage into the risk assessment model and output the comprehensive risk assessment results.

[0193] The early warning level determination module 26 is used to compare the comprehensive risk assessment results with the preset risk threshold range to determine the current early warning level;

[0194] The control command execution module 27 is used to generate a corresponding generation control command based on the warning level and send it to the relevant execution mechanism.

[0195] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 2 The functions of the following modules are shown: visual perception module 21, structural data acquisition module 22, feature data fusion module 23, structural data prediction module 24, comprehensive risk assessment module 25, early warning level determination module 26, and control command execution module 27.

[0196] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0197] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0198] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0199] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in any embodiment of the real-time monitoring method for construction hanging basket operation risk provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0200] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0201] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0202] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0203] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0204] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0205] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0206] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0207] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for real-time monitoring of operational risks of construction hanging baskets, characterized in that, include: Video monitoring data is acquired through a camera, and the video monitoring data is analyzed based on computer vision technology to extract the working condition information and the first time-series feature set of the construction hanging basket. The structural state data of the construction formwork is acquired through sensors, and its features are extracted to obtain a second time-series feature set; which also includes: The monitoring data operated under various extreme environmental conditions were used to construct a validation dataset; the extreme environmental conditions included heavy fog, rainstorms, dust storms, and physical obstructions that could degrade the quality of the video monitoring data. The monitoring data is analyzed using computer vision technology, and features are extracted to obtain a third time-series feature set; Based on the validation dataset, the success rate of extracting the third temporal feature set was evaluated, and the performance degradation curve of computer vision technology under different environmental conditions was quantified. Based on the performance degradation curve, the fusion weights of the first time-series feature set and the second time-series feature set are adjusted, including: The confidence level of computer vision technology under the current environmental conditions is determined based on the performance degradation curve; the weight coefficients of the first time-series feature set are calculated based on the confidence level; the complementary weight coefficients of the second time-series feature set are calculated based on the weight coefficients and the preset baseline weights of the sensor data; a weighted fusion algorithm is used to fuse the first time-series feature set according to the weight coefficients and the second time-series feature set according to the complementary weight coefficients; when the confidence level is less than the preset environmental adaptation threshold, the sensor data-dominated mode is activated, in which the weight of the second time-series feature set is greater than that of the first time-series feature set; The first time-series feature set, the second time-series feature set, and the operating condition information are fused to generate a fused feature vector. The fused feature vector is input into a deep learning-based hanging basket state prediction model to obtain the future change trajectory of the key parameters of the construction hanging basket and the predicted load distribution for the next construction stage; wherein, The fused feature vector is input into the encoder of the hanging basket state prediction model, and its temporal features are extracted through a multi-layer gated loop unit to obtain the encoded state sequence. The encoded state sequence is input into the decoder of the hanging basket state prediction model to obtain prediction values ​​for multiple future time steps; the prediction values ​​include instantaneous prediction values ​​of the key parameters and instantaneous prediction values ​​of the load distribution. The instantaneous predicted values ​​of the key parameters are sequentially integrated to generate the future change trajectory of the key parameters of the construction hanging basket; The instantaneous predicted values ​​of the load distribution are sequentially integrated to generate the predicted load distribution for the next construction stage; The fused feature vector, the future change trajectory, and the predicted load distribution for the next construction stage are input into the risk assessment model, and a comprehensive risk assessment result is output. The comprehensive risk assessment results are compared with the preset risk threshold range to determine the current warning level; Based on the warning level, a corresponding generation control command is generated and sent to the relevant execution agency.

2. The method for real-time monitoring of operational risks of construction hanging baskets according to claim 1, characterized in that, The analysis of the video monitoring data based on computer vision technology, and the extraction of the working condition information and the first time-series feature set of the construction hanging basket, includes: The video monitoring data is subjected to frame sampling processing to obtain multiple consecutive video frames; The video frames are identified to obtain the key components of the construction hanging basket and their pixel-level contour information is acquired. Based on the contour information and the camera calibration parameters, the spatial position and attitude angle of the key component are calculated through spatial geometric transformation. Based on a multi-target tracking algorithm, the spatial position and the attitude angle are correlated in a time sequence to obtain the motion trajectory of the key component; The motion trajectory is analyzed to obtain the motion state and deformation trend of the construction formwork; The motion state, the deformation trend, the spatial position, and the attitude angle are used as the first temporal feature set; The motion state and deformation trend are matched with a preset working condition discrimination rule library to obtain the working condition information of the construction formwork. The working condition information includes static pouring working conditions and dynamic moving working conditions.

3. The method for real-time monitoring of operational risks of construction hanging baskets according to claim 2, characterized in that, The step of performing a time-series correlation between the spatial position and the attitude angle based on a multi-target tracking algorithm to obtain the motion trajectory of the key component includes: Kalman filtering is used to predict the position and state of the key component in the next frame; The predicted location is temporally correlated with the location of key components in the current frame using the Hungarian algorithm to obtain the correlation result. Based on the correlation results, the motion state of the key component is updated, and the continuous position coordinates of the key component in the time series are generated to obtain the motion trajectory of the key component.

4. The method for real-time monitoring of operational risks of construction hanging baskets according to claim 3, characterized in that, The step of temporally associating the predicted location with the location of key components in the current frame using the Hungarian algorithm to obtain the association result includes: Calculate the intersection-union ratio, center point distance, and shape similarity between the predicted location and the location of key components in the current frame, and construct them as a cost matrix; The cost matrix is ​​normalized to obtain the standard cost matrix; Matches in the standard cost matrix with an intersection-union ratio less than a preset first threshold are marked as invalid matches; The Hungarian algorithm is used to solve for the optimal matching of the processed standard cost matrix to obtain matching pairs. Based on the matching pairs and the invalid matching items, distinguish the key components that have been successfully associated, the newly emerging key components, and the key components that failed to be matched, and obtain the association results.

5. The method for real-time monitoring of operational risks of construction hanging baskets according to claim 3, characterized in that, Before updating the motion state of the key component based on the correlation result, generating the continuous position coordinates of the key component in the time series, and obtaining the motion trajectory of the key component, the method further includes: For a key component that is successfully matched, its state and motion trajectory are updated using its position in the current frame. For critical components that fail to match, if they are detected in multiple consecutive frames and the confidence level is greater than the preset confidence threshold, they are identified as newly emerging critical components and their motion trajectories are created. If a critical component fails to match successfully within a preset first time window, it is determined that the critical component has left the monitoring area, and tracking of it is terminated.

6. The method for real-time monitoring of operational risks of construction hanging baskets according to claim 4, characterized in that, After marking matches in the standard cost matrix whose intersection-union ratio is less than a preset first threshold as invalid matches, the method further includes: Acquire auxiliary sensor data synchronized with the video monitoring data, including laser ranging point cloud or millimeter-wave radar data; The predicted location of the key component corresponding to the invalid match is spatially verified with the auxiliary sensor data. If the spatial location verification passes, the corresponding invalid match is corrected to a valid match and reintroduced into the standard cost matrix; If the spatial location verification fails, it remains an invalid match and its corresponding key component is marked as a suspected interference source.

7. The method for real-time monitoring of operational risks of construction hanging baskets according to claim 6, characterized in that, If the spatial location verification fails, it remains an invalid match, and the critical component is marked as a suspected interference source, including: Within a preset time window, the suspected interference sources are re-identified and tracked; If, within the time window, the number of times the suspected interference source is successfully matched in consecutive video frames is greater than a preset second threshold, then its suspected interference source label is removed, and it is used as a newly emerging key component to create a new motion trajectory. If the number of successful matches of the suspected interference source is still less than the second threshold when the time window ends, it will be determined to be an interference source.

8. The method for real-time monitoring of operational risks of construction hanging baskets according to claim 1, characterized in that, The step of inputting the encoded state sequence into the decoder of the hanging basket state prediction model to obtain predicted values ​​for multiple future time steps includes: At each prediction time step of the decoder, the association weight between the current decoding time step and the historical encoding time step is calculated through a temporal attention mechanism to obtain the association weight; Based on the association weights, all hidden states of the encoder are weighted and summed to obtain the encoding state sequence at the current time step; The encoded state sequence of the current time step is concatenated with the output of the decoder at the previous time step. The concatenated features are nonlinearly transformed by a fully connected layer to output the predicted value at the current time step. When the prediction of a preset future time step is reached, the predicted values ​​of the multiple future time steps are obtained.

9. A real-time monitoring system for the operational risks of construction hanging baskets, characterized in that, include: The working condition visual perception module is used to acquire video monitoring data through a camera, analyze the video monitoring data based on computer vision technology, and extract the working condition information and the first time-series feature set of the construction hanging basket. A structural data acquisition module is used to acquire structural state data of the construction formwork through sensors and extract its features to obtain a second time-series feature set; it also includes: The monitoring data operated under various extreme environmental conditions were used to construct a validation dataset; the extreme environmental conditions included heavy fog, rainstorms, dust storms, and physical obstructions that could degrade the quality of the video monitoring data. The monitoring data is analyzed using computer vision technology, and features are extracted to obtain a third time-series feature set; Based on the validation dataset, the success rate of extracting the third temporal feature set was evaluated, and the performance degradation curve of computer vision technology under different environmental conditions was quantified. Based on the performance degradation curve, the fusion weights of the first time-series feature set and the second time-series feature set are adjusted, including: The confidence level of computer vision technology under the current environmental conditions is determined based on the performance degradation curve; the weight coefficients of the first time-series feature set are calculated based on the confidence level; the complementary weight coefficients of the second time-series feature set are calculated based on the weight coefficients and the preset baseline weights of the sensor data; a weighted fusion algorithm is used to fuse the first time-series feature set according to the weight coefficients and the second time-series feature set according to the complementary weight coefficients; when the confidence level is less than the preset environmental adaptation threshold, the sensor data-dominated mode is activated, in which the weight of the second time-series feature set is greater than that of the first time-series feature set; The feature data fusion module is used to fuse the first time-series feature set, the second time-series feature set and the operating condition information to generate a fused feature vector. The structural data prediction module is used to input the fused feature vector into a deep learning-based hanging basket state prediction model to obtain the future change trajectory of the key parameters of the construction hanging basket and the predicted load distribution for the next construction stage; wherein, The fused feature vector is input into the encoder of the hanging basket state prediction model, and its temporal features are extracted through a multi-layer gated loop unit to obtain the encoded state sequence. The encoded state sequence is input into the decoder of the hanging basket state prediction model to obtain prediction values ​​for multiple future time steps; the prediction values ​​include instantaneous prediction values ​​of the key parameters and instantaneous prediction values ​​of the load distribution. The instantaneous predicted values ​​of the key parameters are sequentially integrated to generate the future change trajectory of the key parameters of the construction hanging basket; The instantaneous predicted values ​​of the load distribution are sequentially integrated to generate the predicted load distribution for the next construction stage; The comprehensive risk assessment module is used to input the fused feature vector, the future change trajectory, and the predicted load distribution of the next construction stage into the risk assessment model, and output the comprehensive risk assessment result. The early warning level determination module is used to compare the comprehensive risk assessment results with a preset risk threshold range to determine the current early warning level; The control command execution module is used to generate a corresponding generation control command based on the warning level and send it to the relevant execution mechanism.

Images