Operation area construction behavior intelligent identification method based on attention mechanism

Abstract

The invention discloses an attention mechanism-based operation area construction behavior intelligent identification method, which comprises the following steps of S1, collecting and preprocessing multi-modal data, and generating a standardized construction action time sequence data set; s2, constructing a Transform network model, and extracting spatial features and time features; s3, the Transform network model is optimized based on a parrot optimization algorithm, and optimal hyper-parameter configuration is obtained; s4, carrying out training by utilizing the optimized Transform network model, and carrying out supervised learning by adopting a cross entropy loss function; s5, performing construction behavior intelligent identification, and outputting a construction action category and an identification confidence coefficient; and S6, triggering a real-time early warning mechanism, and pushing early warning information to a construction safety management platform. According to the method, efficient and accurate construction behavior recognition and high-risk action detection are realized by optimizing the Transform network model and dynamically adjusting the hyper-parameters, and the safety and management efficiency of a construction site are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an attention mechanism-based intelligent recognition method for construction behavior in a work area. Background Art

[0002] With the continuous expansion of construction projects, infrastructure development, and industrial projects, the work environment at construction sites is becoming increasingly complex. With a large number of construction workers and heavy workloads, the difficulty of safety management has increased significantly. Traditional construction safety supervision relies primarily on manual inspections and post-analysis through playback of surveillance camera footage. This approach not only suffers from poor real-time performance, delayed response, and high dependence on personnel, but also, due to fatigue and subjective biases in manual observation, it can easily lead to missed detections and misjudgments of safety hazards. With the development of intelligent technology, intelligent recognition technology for construction behavior has gradually been proposed. It aims to use automated means to analyze construction worker behavior in real time, promptly detect high-risk actions, provide early warnings, reduce accident rates, and improve the overall safety level of the work area.

[0003] Existing methods for identifying construction actions primarily rely on traditional computer vision algorithms or early shallow neural network models. Most of these methods rely on manual feature extraction and rule matching, using methods such as motion trajectories, changes in movement amplitude, or classification rules based on preset thresholds to roughly identify construction actions. However, due to the limitations of manual features and the inflexibility of rule definition, these methods generally exhibit low recognition accuracy in real-world scenarios such as complex environmental changes, diverse construction actions, and frequent occlusion and lighting changes. Furthermore, their ability to detect high-risk actions such as illegal climbing and working without safety equipment is extremely limited.

[0004] With the development of deep learning technologies, particularly convolutional neural networks (CNNs) and recurrent neural networks (RNNs), researchers have begun to apply these networks to construction action recognition tasks. CNNs are highly capable of extracting spatial features from images or video frames, while RNNs, particularly long-short-term memory (LSTM) networks, offer advantages in time series modeling. Therefore, the CNN-LSTM-based action recognition framework became the primary technical approach for early intelligent construction recognition systems. However, these CNN-RNN-based architectures still have significant drawbacks. Convolutional neural networks are effective for extracting local features but lack the ability to model global dependencies. Furthermore, recurrent neural networks suffer from long training times, vanishing gradients, and poor ability to model long-range dependencies. These issues lead to significant performance degradation during complex action recognition over extended periods, making them difficult to meet the real-time recognition requirements for continuous and dynamic construction actions.

[0005] In recent years, with the successful application of the Transformer model in natural language processing and computer vision, its attention-based features enable the model to dynamically focus on different parts of the input data globally, significantly improving its ability to model spatiotemporal features. Consequently, some research has incorporated the Transformer architecture for action recognition. Through its self-attention mechanism, the Transformer can simultaneously model both the spatial variations and temporal evolution of construction actions, demonstrating superior performance compared to CNN-RNN models in complex construction scenarios. However, current Transformer-based methods for construction action recognition mostly use a generic architecture and lack specialized optimizations tailored to the specific characteristics of construction scenarios, such as the sparse nature of high-risk actions, the sudden nature of abnormal behaviors, and the high demand for real-time response. Furthermore, existing methods often overlook the significant impact of model hyperparameter configuration on recognition performance, such as position encoding strategies, feedforward network expansion, the number of heads in the attention mechanism, and the selection criteria. This results in limited adaptability to diverse construction scenarios and significant fluctuations in recognition performance.

[0006] In addition, in existing intelligent construction recognition systems, most methods only focus on the overall classification accuracy of construction actions, ignoring the specialized recognition and priority detection mechanisms for high-risk actions. High-risk construction actions such as aerial work, live-line work, and heavy object handling are often extremely unbalanced in data distribution compared to general work actions. Traditional recognition methods are prone to insufficient recognition of high-risk actions, resulting in hidden dangers of missed detection, and are unable to meet the actual needs of real-time early warning for construction safety. When dealing with high-risk actions, existing Transformer-based behavior recognition methods often suffer from performance degradation due to the inability of the attention mechanism to effectively and adaptively highlight the characteristics of high-risk actions. At the same time, some existing systems ignore the strong real-time requirements of the recognition system in construction scenarios. The inference delay is too large, and real-time feedback cannot be achieved, which increases safety risks.

[0007] Therefore, how to provide an intelligent recognition method for construction behavior in the work area based on the attention mechanism is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0008] One purpose of the present invention is to propose a method for intelligent identification of construction behaviors in work areas based on an attention mechanism. The present invention makes full use of the spatial and temporal feature modeling capabilities of the Transformer network, combines the Parrot optimization algorithm to dynamically optimize the hyperparameters of the Transformer network, and describes in detail the technical process of using multimodal perception data to realize construction behavior classification and intelligent identification of high-risk actions. The method has the advantages of high recognition accuracy, sensitivity to high-risk action detection, good real-time performance, and strong ability to adapt to complex construction environments.

[0009] According to an embodiment of the present invention, a method for intelligently identifying construction behaviors in a work area based on an attention mechanism includes the following steps: S1. Collect multimodal data of construction workers in the work area and preprocess the multimodal data to generate a standardized construction action time series dataset; S2. Constructing a Transformer network model to extract spatial and temporal features based on the standardized construction action time series dataset; S3. Optimize the hyperparameters of the Transformer network model based on the Parrot optimization algorithm to obtain the optimal hyperparameter configuration, and dynamically adjust the Transformer network structure based on the optimal hyperparameter configuration. At the same time, introduce a local attention and global attention fusion mechanism to adaptively adjust the attention weight distribution strategy; S4. Use the optimized Transformer network model to train the standardized construction action time series dataset and use the cross entropy loss function for supervised learning; S5. Apply the trained optimized Transformer network model to the real-time data stream input of the construction work area to perform intelligent identification of construction behaviors and output specific construction action categories and recognition confidence levels. S6. When high-risk or illegal construction behavior is detected, a real-time early warning mechanism is triggered based on preset safety rules, and the early warning information is pushed to the construction safety management platform.

[0010] Optionally, the multimodal data specifically includes RGB video images, depth images and skeletal key point coordinate information, which are used to comprehensively characterize the spatial posture changes and motion behavior characteristics of construction workers in the work area.

[0011] Optionally, the preprocessing of the multimodal data specifically includes denoising, normalization and time series synchronization, so as to improve the accuracy of construction action feature extraction.

[0012] Optionally, the S2 specifically includes: S21. Represent the standardized construction action time series dataset as an input sequence ,in Indicates that at time step The collected multimodal feature vectors, , is the sequence length; S22, input sequence Perform position coding and adopt dynamic position coding fusion mechanism, which generates the final position coding according to the following formula : ; in, is the position encoding fusion ratio, is the sinusoidal position code, Encoding for learnable positions; S23, add the input sequence with the final position encoding Input to the encoder module of the Transformer network model, the encoder module includes a multi-head self-attention mechanism module and a feedforward neural network module; S24. In the multi-head self-attention mechanism module, the number of attention heads involved in the calculation is dynamically adjusted according to the attention head importance gating mechanism, specifically including: obtaining the importance score of each attention head , the retention score is greater than or equal to the attention head importance threshold The attention head participates in feature modeling; S25. In the feedforward neural network module, based on the dynamic feedforward expansion factor Determine the hidden layer expansion dimension ; S26. Normalize the feature representation after processing by the multi-head self-attention mechanism and the feedforward neural network module, and output a sequence of joint representations of spatial features and temporal features.

[0013] Optionally, the S3 specifically includes: S31, initialize the parrot optimization algorithm population, set the population size to , each parrot individual corresponds to a set of Transformer network model hyperparameters, which include the position encoding fusion ratio , feedforward expansion factor and attention head importance threshold ; S32. Evaluate each parrot individual based on the comprehensive fitness function The fitness value of , comprehensive fitness function Defined as: ; in, is the accuracy of construction action recognition, For high-risk action detection accuracy, is the average attention entropy, is the inference delay time, 、 、 、 is a positive weight coefficient; S33, sort the population according to the fitness value, and sort the top Parrot individuals are divided into advanced layers, and the remaining parrot individuals are divided into ordinary layers, among which is the advanced layer scaling factor; S34: For the individual parrots in the high-level layer, perform local imitation learning operations and update the individual parrot positions to : ; in, For the The position of the parrot with the best fitness in the generation, is a random number; S35. Randomly select high-level parrot individuals for ordinary parrot individuals Perform imitation learning based on imitation credibility weight Update the parrot individual position to : ; in, To simulate the credibility weight, it is defined as: ; in, For the parrot individual being imitated The construction action recognition accuracy of the generation, To prevent small positive numbers with zero denominators, For the parrot individual being imitated The construction action recognition accuracy of the generation, For the Daidi The accuracy of identifying construction actions of individual parrots in a population, For the Daidi The accuracy of identifying construction actions of individual parrots in a population; S36. After completing the imitation learning, the dynamic migration expansion mechanism is introduced to apply global perturbations to each parrot individual and update the position to : ; in, and are the positions of two different parrot individuals randomly selected from the population, and are the maximum and minimum migration radius, is the maximum iteration number; S37. Define the final update result of each parrot individual as , and form the The new population of the generation; S38. Recalculate the fitness value of the parrot individuals in the updated population , update the current optimal fitness parrot individual ; S39, determine whether the termination condition is met, if the current iteration algebra reaches the maximum iteration algebra Or if the fitness change is less than the set convergence threshold, the iteration is terminated, otherwise it returns to step S33 to continue the iteration; S310, extract the hyperparameter configuration corresponding to the optimal parrot individual in the final population, that is, the optimal position encoding fusion ratio , optimal feedforward expansion factor and the optimal attention head importance threshold , and dynamically adjust the Transformer network structure based on the optimal hyperparameter configuration, complete the position encoding fusion adjustment, feedforward neural network expansion and attention head screening setting, which are used for construction behavior intelligent recognition training and reasoning.

[0014] Optionally, the S4 specifically includes: S41. Based on the determined optimal positional encoding fusion ratio, feedforward expansion factor, and attention head importance threshold, an optimized Transformer network model structure is constructed. S42, dividing the standardized construction action time series data set into a training data subset and a verification data subset according to a preset ratio of the training set to the verification set; S43, batch-dividing the training data subsets, setting batch size, learning rate, and optimizer type training hyperparameters, and initializing the optimized Transformer network model weight parameters; S44, using the cross entropy loss function as the training target, inputting the training data subset into the optimized Transformer network model, performing forward propagation to calculate the output of the optimized Transformer network model, and calculating the loss value based on the output result and the true label; S45. Calculate the gradient information of the loss function with respect to the model parameters based on the back-propagation algorithm, and use the optimizer to iteratively update the Transformer network model parameters. S46. After each training cycle, the performance indicators of the optimized Transformer network model in the construction action recognition task are evaluated based on the validation data subset, and the construction action classification accuracy and loss value change trend of the optimized Transformer network model on the validation set are recorded until the training is completed.

[0015] Optionally, the S5 specifically includes: S51. Deploy the trained optimized Transformer network model in the construction behavior intelligent recognition system in the work area; S52, using smart cameras and depth camera sensing devices to collect video streams, depth images, and skeleton key point data of the construction site in real time, and perform synchronous processing and standardized preprocessing on the collected data; S53, organizing the pre-processed real-time data stream of the construction site into a construction action time sequence input sequence in chronological order, and inputting it into the optimized Transformer network model for feature extraction and time sequence modeling; S54, inferring the input data through the optimized Transformer network model, and outputting the prediction result of the construction action category corresponding to each time step; S55. Based on the classification results output by the optimized Transformer network model, the corresponding recognition confidence is calculated simultaneously. The confidence value is used to quantify the credibility of the prediction results of each construction action category; S56. Format the construction action categories and corresponding confidence results, and output them to the construction safety monitoring platform in real time for construction behavior monitoring, abnormal behavior warning, and behavior log recording.

[0016] Optionally, the intelligent identification system for construction behavior in the work area refers to a system based on multimodal sensing equipment and an optimized Transformer network model to collect construction personnel motion data in real time, classify construction behaviors and identify high-risk behaviors, and use them for intelligent monitoring and safety warning management of the construction site.

[0017] The beneficial effects of the present invention are: This paper proposes an attention-based intelligent identification method for construction activities in work areas. This method overcomes the significant shortcomings of existing construction activity monitoring technologies in terms of accuracy, high-risk action detection, real-time model responsiveness, and adaptability to complex environments. By constructing an optimized Transformer network structure and leveraging the attention mechanism to globally model construction action characteristics in both spatial and temporal dimensions, the model's ability to capture the spatiotemporal variations of construction actions is effectively enhanced. This enables more accurate action classification and state recognition, particularly in construction sites characterized by high variability, frequent occlusion, and complex continuous motion.

[0018] This paper uses the Parrot optimization algorithm to dynamically and jointly optimize the key hyperparameters of the Transformer network. It introduces a positional encoding fusion mechanism, a dynamic adjustment mechanism for the feedforward expansion factor, and an attention head importance screening mechanism, enabling the Transformer network structure to be adaptively adjusted according to the characteristics of construction actions, thereby improving the model's recognition stability and generalization capabilities in different construction scenarios. At the same time, by designing a comprehensive fitness function, the overall construction action recognition accuracy, high-risk action detection accuracy, attention concentration, and inference latency are comprehensively considered to ensure that the optimization process not only focuses on conventional classification performance, but also takes into account the timely detection capability of high-risk actions and the real-time response requirements of the system, thereby significantly enhancing the effectiveness and reliability of the construction intelligent recognition system in actual application scenarios.

[0019] Furthermore, during the intelligent identification of construction actions, this invention establishes a real-time intelligent feedback mechanism by outputting the construction action category and corresponding recognition confidence. This mechanism dynamically adjusts the response strategy based on the confidence level, enhancing the credibility of identification decisions and the safety of system operation. The prioritized detection and immediate warning mechanism for high-risk construction actions further enhances construction safety management, reduces the risk of construction accidents, and improves overall construction efficiency and intelligent management of the work area.

[0020] In summary, the present invention not only achieves a breakthrough in the accuracy of construction behavior recognition, but also has obvious advantages in terms of high-risk action detection sensitivity, model real-time performance and adaptability to the construction environment. It can effectively make up for the many shortcomings of existing technologies in the application of intelligent identification on construction sites, and has good practical application value and broad promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a flowchart of an intelligent identification method for construction behavior in a work area based on an attention mechanism proposed by the present invention; Figure 2 This is a schematic diagram of the optimization process of dynamically adjusting the Transformer network hyperparameters based on the Parrot optimization algorithm for an intelligent recognition method of construction behavior in a work area based on an attention mechanism proposed in the present invention. DETAILED DESCRIPTION

[0022] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0023] refer to Figure 1 and Figure 2 , an intelligent recognition method for construction behavior in a work area based on an attention mechanism, comprising the following steps: S1. Collect multimodal data of construction workers in the work area and preprocess the multimodal data to generate a standardized construction action time series dataset; S2. Constructing a Transformer network model to extract spatial and temporal features based on the standardized construction action time series dataset; S3. Optimize the hyperparameters of the Transformer network model based on the Parrot optimization algorithm to obtain the optimal hyperparameter configuration, and dynamically adjust the Transformer network structure based on the optimal hyperparameter configuration. At the same time, introduce a local attention and global attention fusion mechanism to adaptively adjust the attention weight distribution strategy; S4. Use the optimized Transformer network model to train the standardized construction action time series dataset and use the cross entropy loss function for supervised learning; S5. Apply the trained optimized Transformer network model to the real-time data stream input of the construction work area to perform intelligent identification of construction behaviors and output specific construction action categories and recognition confidence levels. S6. When high-risk or illegal construction behavior is detected, a real-time early warning mechanism is triggered based on preset safety rules, and the early warning information is pushed to the construction safety management platform.

[0024] In this embodiment, the multimodal data specifically includes RGB video images, depth images and skeleton key point coordinate information, which are used to comprehensively characterize the spatial posture changes and motion behavior characteristics of construction workers in the work area.

[0025] In this embodiment, the preprocessing of the multimodal data specifically includes noise removal, normalization and time series synchronization, which is used to improve the accuracy of construction action feature extraction.

[0026] In this embodiment, S2 specifically includes: S21. Represent the standardized construction action time series dataset as an input sequence ,in Indicates that at time step The collected multimodal feature vectors, , is the sequence length; S22, input sequence Perform position coding and adopt dynamic position coding fusion mechanism, which generates the final position coding according to the following formula : ; in, is the position encoding fusion ratio, is the sinusoidal position code, Encoding for learnable positions; S23, add the input sequence with the final position encoding Input to the encoder module of the Transformer network model, the encoder module includes a multi-head self-attention mechanism module and a feedforward neural network module; S24. In the multi-head self-attention mechanism module, the number of attention heads involved in the calculation is dynamically adjusted according to the attention head importance gating mechanism, specifically including: obtaining the importance score of each attention head , the retention score is greater than or equal to the attention head importance threshold The attention head participates in feature modeling; S25. In the feedforward neural network module, based on the dynamic feedforward expansion factor Determine the hidden layer expansion dimension ; S26. Normalize the feature representation after processing by the multi-head self-attention mechanism and the feedforward neural network module, and output a sequence of joint representations of spatial features and temporal features.

[0027] In this embodiment, S3 specifically includes: S31, initialize the parrot optimization algorithm population, set the population size to , each parrot individual corresponds to a set of Transformer network model hyperparameters, which include the position encoding fusion ratio , feedforward expansion factor and attention head importance threshold ; S32. Evaluate each parrot individual based on the comprehensive fitness function The fitness value of , comprehensive fitness function Defined as: ; in, is the accuracy of construction action recognition, For high-risk action detection accuracy, is the average attention entropy, is the inference delay time, 、 、 、 is a positive weight coefficient; S33, sort the population according to the fitness value, and sort the top Parrot individuals are divided into advanced layers, and the remaining parrot individuals are divided into ordinary layers, among which is the advanced layer scaling factor; S34: For the individual parrots in the high-level layer, perform local imitation learning operations and update the individual parrot positions to : ; in, For the The position of the parrot with the best fitness in the generation, is a random number; S35. Randomly select high-level parrot individuals for ordinary parrot individuals Perform imitation learning based on imitation credibility weight Update the parrot individual position to : ; in, To simulate the credibility weight, it is defined as: ; in, For the parrot individual being imitated The construction action recognition accuracy of the generation, To prevent small positive numbers with zero denominators, For the parrot individual being imitated The construction action recognition accuracy of the generation, For the Daidi The accuracy of identifying construction actions of individual parrots in a population, For the Daidi The accuracy of identifying construction actions of individual parrots in a population; S36. After completing the imitation learning, the dynamic migration expansion mechanism is introduced to apply global perturbations to each parrot individual and update the position to : ; in, and are the positions of two different parrot individuals randomly selected from the population, and are the maximum and minimum migration radius, is the maximum iteration number; S37. Define the final update result of each parrot individual as , and form the The new population of the generation; S38. Recalculate the fitness value of the parrot individuals in the updated population , update the current optimal fitness parrot individual ; S39, determine whether the termination condition is met, if the current iteration algebra reaches the maximum iteration algebra Or if the fitness change is less than the set convergence threshold, the iteration is terminated, otherwise it returns to step S33 to continue the iteration; S310, extract the hyperparameter configuration corresponding to the optimal parrot individual in the final population, that is, the optimal position encoding fusion ratio , optimal feedforward expansion factor and the optimal attention head importance threshold , and dynamically adjust the Transformer network structure based on the optimal hyperparameter configuration, complete the position encoding fusion adjustment, feedforward neural network expansion and attention head screening setting, which are used for construction behavior intelligent recognition training and reasoning.

[0028] In this embodiment, the S4 specifically includes: S41. Based on the determined optimal positional encoding fusion ratio, feedforward expansion factor, and attention head importance threshold, an optimized Transformer network model structure is constructed. S42, dividing the standardized construction action time series data set into a training data subset and a verification data subset according to a preset ratio of the training set to the verification set; S43, batch-dividing the training data subsets, setting batch size, learning rate, and optimizer type training hyperparameters, and initializing the optimized Transformer network model weight parameters; S44, using the cross entropy loss function as the training target, inputting the training data subset into the optimized Transformer network model, performing forward propagation to calculate the output of the optimized Transformer network model, and calculating the loss value based on the output result and the true label; S45. Calculate the gradient information of the loss function with respect to the model parameters based on the back-propagation algorithm, and use the optimizer to iteratively update the Transformer network model parameters. S46. After each training cycle, the performance indicators of the optimized Transformer network model in the construction action recognition task are evaluated based on the validation data subset, and the construction action classification accuracy and loss value change trend of the optimized Transformer network model on the validation set are recorded until the training is completed.

[0029] In this embodiment, the S5 specifically includes: S51. Deploy the trained optimized Transformer network model in the construction behavior intelligent recognition system in the work area; S52, using smart cameras and depth camera sensing devices to collect video streams, depth images, and skeleton key point data of the construction site in real time, and perform synchronous processing and standardized preprocessing on the collected data; S53, organizing the pre-processed real-time data stream of the construction site into a construction action time sequence input sequence in chronological order, and inputting it into the optimized Transformer network model for feature extraction and time sequence modeling; S54, inferring the input data through the optimized Transformer network model, and outputting the prediction result of the construction action category corresponding to each time step; S55. Based on the classification results output by the optimized Transformer network model, the corresponding recognition confidence is calculated simultaneously. The confidence value is used to quantify the credibility of the prediction results of each construction action category; S56. Format the construction action categories and corresponding confidence results, and output them to the construction safety monitoring platform in real time for construction behavior monitoring, abnormal behavior warning, and behavior log recording.

[0030] In this embodiment, the intelligent identification system for construction behavior in the work area refers to a system based on multimodal sensing equipment and an optimized Transformer network model to collect construction personnel motion data in real time, classify construction behaviors and identify high-risk behaviors, and use them for intelligent monitoring and safety warning management of the construction site.

[0031] Example 1: To verify the feasibility of this invention, it was applied to the work area of a large-scale construction project. During the construction process, this construction site required a large number of complex and high-risk activities such as high-altitude work, lifting and handling, and welding. The large number of workers and the heavy workload made safety management difficult. To effectively improve safety management efficiency and reduce safety accidents caused by improper operation by construction workers, the site decided to introduce an intelligent construction behavior recognition system based on an attention mechanism. This system relies on real-time data acquisition equipment, such as smart cameras, depth cameras, and motion capture sensors, to monitor workers' behavior in real time and detect high-risk actions through an optimized Transformer network model.

[0032] At the construction site, all workers wear sensor-equipped helmets and work clothes with motion capture capabilities. Multiple high-precision smart cameras are installed. These devices collect real-time dynamic data about the workers, including their movements, positions, speed, and posture. This information is then transmitted to the intelligent construction behavior recognition system for processing. The system analyzes this multimodal data (video streams, motion data, depth images, etc.) using an optimized Transformer network model.

[0033] First, the system normalizes the collected real-time data stream and feeds it into a trained Transformer network model. Using an attention mechanism, the model extracts spatial and temporal characteristics of workers and classifies construction actions. Through attention head screening and adaptive adjustment strategies, the system can focus on identifying high-risk actions, such as those involving high-altitude work and live-line work. During this process, the Parrot optimization algorithm dynamically optimizes the model's hyperparameters, ensuring its adaptability to complex construction environments and ever-changing work scenarios.

[0034] When the system identifies construction workers engaging in dangerous behavior, such as not wearing safety belts or standing at the edge of an elevated work area, it immediately triggers an early warning mechanism and displays the specific type of dangerous action and corresponding warning information on the safety monitoring platform. Workers can then intervene in real time based on the system's warnings to prevent accidents. This system not only improves safety during construction but also enhances the efficiency of construction management by reducing the workload of manual inspections.

[0035] Table 1 Comparison of application effects of construction behavior intelligent recognition system ;

[0036] By comparing the application effects of the manual inspection system and the construction behavior intelligent recognition system of the present invention, we can clearly see the significant advantages of the intelligent recognition system in construction safety management.

[0037] First, the missed detection rate metric reveals the difference between the two systems in their ability to identify high-risk behaviors. The manual inspection system has a high missed detection rate of 18%, meaning that approximately 18% of high-risk behaviors are missed during each hour of inspection, posing a significant safety hazard. The intelligent recognition system, on the other hand, significantly reduces this missed detection rate to 2%, effectively improving the recognition rate of high-risk behaviors on construction sites and reducing potential safety risks.

[0038] Secondly, recognition accuracy is a key indicator of a system's classification capabilities. In manual inspection systems, recognition accuracy is only 82%, meaning 18% of recognition results are incorrect, potentially leading to erroneous safety decisions. However, the intelligent recognition system excels in this regard, achieving an accuracy rate of 98%. This significantly improved accuracy demonstrates that the intelligent recognition system can more accurately identify construction worker behavior, particularly high-risk behavior, providing more reliable data support for safety management.

[0039] Another key metric is average response time, reflecting the system's delay from identification to response. Traditional manual inspections have an average response time of 15 minutes, a delay that can result in safety hazards not being discovered and addressed promptly, increasing the risk of accidents. In contrast, the intelligent recognition system has an average response time of just 2 minutes, enabling it to react in record time, providing on-site managers with timely safety warnings and significantly improving construction site safety.

[0040] The accuracy of high-risk behavior warnings and the system's false alarm rate further demonstrate the advantages of the intelligent recognition system in detecting high-risk behaviors. The manual patrol system's high-risk behavior warning accuracy was only 65%, while the intelligent recognition system achieved 95%. This demonstrates that the intelligent recognition system can more accurately detect and warn of high-risk behaviors, avoiding the misjudgments and missed detections that can occur with traditional systems. Furthermore, the intelligent recognition system's false alarm rate was only 2%, far lower than the 10% rate of the manual patrol system. This means that while the intelligent recognition system accurately identifies dangerous behaviors, it also avoids excessive false alarms, reducing unnecessary disruptions and workload for management personnel caused by these false alarms.

[0041] Finally, the high-risk operation detection success rate demonstrates the system's ability to identify high-risk actions. While the manual inspection system achieved a success rate of 70%, the intelligent recognition system achieved a success rate of 98%. This demonstrates the intelligent recognition system's robust ability to detect high-risk actions, enabling it to promptly identify dangerous actions on-site, effectively preventing potential safety incidents.

[0042] In summary, the intelligent recognition system significantly outperforms traditional manual inspection systems across multiple key metrics, particularly in high-risk behavior identification, accuracy, response time, and false alarm rate. The introduction of the intelligent recognition system has significantly improved construction site safety management, effectively reducing safety incidents while significantly improving management efficiency and work area safety. The application of this technical solution has fully demonstrated its broad applicability and feasibility in actual construction sites.

[0043] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for intelligent identification of construction behavior in a work area based on an attention mechanism, characterized in that: The steps include: S1. Collect multimodal data of construction workers in the work area and preprocess the multimodal data to generate a standardized construction action time series dataset; S2. Constructing a Transformer network model to extract spatial and temporal features based on the standardized construction action time series dataset; S3. Optimize the hyperparameters of the Transformer network model based on the Parrot optimization algorithm to obtain the optimal hyperparameter configuration, and dynamically adjust the Transformer network structure based on the optimal hyperparameter configuration. At the same time, introduce a local attention and global attention fusion mechanism to adaptively adjust the attention weight distribution strategy; S4. Use the optimized Transformer network model to train the standardized construction action time series dataset and use the cross entropy loss function for supervised learning; S5. Apply the trained optimized Transformer network model to the real-time data stream input of the construction work area to perform intelligent identification of construction behaviors and output the specific construction action category and recognition confidence level. S6. When high-risk or illegal construction behavior is detected, a real-time early warning mechanism is triggered based on preset safety rules, and the early warning information is pushed to the construction safety management platform.

2. The method for intelligently identifying construction behaviors in a work area based on an attention mechanism according to claim 1 is characterized in that: The multimodal data specifically includes RGB video images, depth images and skeletal key point coordinate information, which is used to comprehensively characterize the spatial posture changes and action behavior characteristics of construction workers in the work area.

3. The method for intelligently identifying construction behaviors in a work area based on an attention mechanism according to claim 1 is characterized in that: The preprocessing of the multimodal data specifically includes noise removal, normalization and time series synchronization, which is used to improve the accuracy of construction action feature extraction.

4. The method for intelligently identifying construction behaviors in a work area based on an attention mechanism according to claim 1 is characterized in that: The S2 specifically includes: S21. Represent the standardized construction action time series dataset as an input sequence ,in Indicates that at time step The collected multimodal feature vectors, , is the sequence length; S22, input sequence Perform position coding and adopt dynamic position coding fusion mechanism, which generates the final position coding according to the following formula : ； in, is the position encoding fusion ratio, is the sinusoidal position code, Encoding for learnable positions; S23, add the input sequence with the final position encoding Input to the encoder module of the Transformer network model, the encoder module includes a multi-head self-attention mechanism module and a feedforward neural network module; S24. In the multi-head self-attention mechanism module, the number of attention heads involved in the calculation is dynamically adjusted according to the attention head importance gating mechanism, specifically including: obtaining the importance score of each attention head , the retention score is greater than or equal to the attention head importance threshold The attention head participates in feature modeling; S25. In the feedforward neural network module, based on the dynamic feedforward expansion factor Determine the hidden layer expansion dimension ; S26. Normalize the feature representation after processing by the multi-head self-attention mechanism and the feedforward neural network module, and output a sequence of joint representations of spatial features and temporal features.

5. The method for intelligently identifying construction behaviors in a work area based on an attention mechanism according to claim 1 is characterized in that: The S3 specifically includes: S31, initialize the parrot optimization algorithm population, set the population size to , each parrot individual corresponds to a set of Transformer network model hyperparameters, which include the position encoding fusion ratio , feedforward expansion factor and attention head importance threshold ; S32. Evaluate each parrot individual based on the comprehensive fitness function The fitness value of , comprehensive fitness function Defined as: ； in, is the accuracy of construction action recognition, For the accuracy of high-risk action detection, is the average attention entropy, is the inference delay time, 、 、 、 is a positive weight coefficient; S33, sort the population according to the fitness value, and sort the top Parrot individuals are divided into advanced layers, and the remaining parrot individuals are divided into ordinary layers, among which is the advanced layer scaling factor; S34: For the individual parrots in the high-level layer, perform local imitation learning operations and update the individual parrot positions to : ； in, For the The position of the parrot with the best fitness in the generation, is a random number; S35. Randomly select high-level parrot individuals for ordinary parrot individuals Perform imitation learning based on imitation credibility weight Update the parrot individual position to : ； in, To simulate the credibility weight, it is defined as: ； in, For the parrot individual being imitated The construction action recognition accuracy of the generation, To prevent small positive numbers with zero denominators, For the parrot individual being imitated The construction action recognition accuracy of the generation, For the Daidi The accuracy of identifying construction actions of individual parrots in a population, For the Daidi The accuracy of identifying construction actions of individual parrots in a population; S36. After completing the imitation learning, the dynamic migration expansion mechanism is introduced to apply global perturbations to each parrot individual and update the position to : ； in, and are the positions of two different parrot individuals randomly selected from the population, and are the maximum and minimum migration radius, is the maximum iteration number; S37. Define the final update result of each parrot individual as , and form the The new population of the generation; S38. Recalculate the fitness value of the parrot individuals in the updated population , update the current optimal fitness parrot individual ; S39, determine whether the termination condition is met, if the current iteration algebra reaches the maximum iteration algebra Or if the fitness change is less than the set convergence threshold, the iteration is terminated, otherwise it returns to step S33 to continue the iteration; S310, extract the hyperparameter configuration corresponding to the optimal parrot individual in the final population, that is, the optimal position encoding fusion ratio , optimal feedforward expansion factor and the optimal attention head importance threshold , and dynamically adjust the Transformer network structure based on the optimal hyperparameter configuration, complete the position encoding fusion adjustment, feedforward neural network expansion and attention head screening setting, which are used for construction behavior intelligent recognition training and reasoning.

6. The method for intelligently identifying construction behaviors in a work area based on an attention mechanism according to claim 1 is characterized in that: The S4 specifically includes: S41. Based on the determined optimal positional encoding fusion ratio, feedforward expansion factor, and attention head importance threshold, an optimized Transformer network model structure is constructed. S42, dividing the standardized construction action time series data set into a training data subset and a verification data subset according to a preset ratio of the training set to the verification set; S43, batch-dividing the training data subsets, setting batch size, learning rate, and optimizer type training hyperparameters, and initializing the optimized Transformer network model weight parameters; S44, using the cross entropy loss function as the training target, inputting the training data subset into the optimized Transformer network model, performing forward propagation to calculate the output of the optimized Transformer network model, and calculating the loss value based on the output result and the true label; S45. Calculate the gradient information of the loss function with respect to the model parameters based on the back-propagation algorithm, and use the optimizer to iteratively update the Transformer network model parameters. S46. After each training cycle, the performance indicators of the optimized Transformer network model in the construction action recognition task are evaluated based on the validation data subset, and the construction action classification accuracy and loss value change trend of the optimized Transformer network model on the validation set are recorded until the training is completed.

7. The method for intelligently identifying construction behaviors in a work area based on an attention mechanism according to claim 1 is characterized in that: The S5 specifically includes: S51. Deploy the trained optimized Transformer network model in the construction behavior intelligent recognition system in the work area; S52, using smart cameras and depth camera sensing devices to collect video streams, depth images, and skeleton key point data of the construction site in real time, and perform synchronous processing and standardized preprocessing on the collected data; S53, organizing the pre-processed real-time data stream of the construction site into a construction action time sequence input sequence in chronological order, and inputting it into the optimized Transformer network model for feature extraction and time sequence modeling; S54, inferring the input data through the optimized Transformer network model, and outputting the prediction result of the construction action category corresponding to each time step; S55. Based on the classification results output by the optimized Transformer network model, the corresponding recognition confidence is calculated simultaneously. The confidence value is used to quantify the credibility of the prediction results of each construction action category; S56. Format the construction action categories and corresponding confidence results, and output them to the construction safety monitoring platform in real time for construction behavior monitoring, abnormal behavior warning, and behavior log recording.

8. The method for intelligently identifying construction behaviors in a work area based on an attention mechanism according to claim 1 is characterized in that: The intelligent identification system for construction behavior in the work area refers to a system based on multimodal sensing equipment and an optimized Transformer network model to collect construction personnel's motion data in real time, classify construction behaviors and identify high-risk behaviors, and is used for intelligent monitoring and safety warning management of the construction site.

Images