A construction worker behavior safety management correction method and system based on AI technology

Abstract

The present invention discloses a method and system for correcting the behavior safety management of construction workers based on AI technology, including a cloud computing center, a regional edge computing center and a front-end sensing device; the platform for the training function of the unsafe behavior recognition model in the cloud computing center imports the video image data set of unsafe behavior instances collected by the front-end sensing device from the construction site, repeatedly trains the unsafe behavior algorithm to obtain the unsafe behavior model, and then directly loads the model to the regional edge computing center, thereby performing artificial intelligence analysis on the video image data set, intelligently identifying the trained unsafe behaviors and classifying and storing them; and performing sample annotation on the identified images, on the one hand, providing them to the artificial intelligence model training platform for training, and on the other hand, providing the required case samples to the platform for the precise intervention function of behavior correction. The system effectively reduces the unsafe behaviors of workers in the production process and strengthens behavioral safety.

Description

Technical Field

[0001] The present invention relates to the field of safety management in the construction industry, and in particular to an AI-based construction worker behavior safety management correction method and system. Background Art

[0002] With the rapid development of the construction industry, safety accidents are also frequent. Unsafe behavior is one of the direct causes of accidents, and the development of safe behavior is one of the important ways to prevent accidents. As the most important method of safety prevention and control management, behavioral safety has been gradually applied in various fields around the world after 40 years of theoretical exploration and practical testing. It mainly focuses on individual behavior, with the goal of reducing unsafe behaviors in production operations and strengthening behavioral safety, so as to prevent accidents and improve safety performance. As an important tool for safety management practice, behavioral safety has been widely used in the construction industry and other fields.

[0003] Taking the behavior safety of construction workers as the research object, the behavior safety method of the construction industry (Behavior-Based Safety in Construction Industry, hereinafter referred to as BBS-CI) was established according to the steps and processes of "defining target behaviors → observing and recording unsafe behaviors → correcting behaviors → feedback incentives" to correct the unsafe behaviors of construction workers, so that construction workers can develop good construction habits and safety awareness, promote the formation of a better safe working environment, and thus improve the safety level of the construction industry. This method is an observation and action method based on behavioral safety. By establishing the difference between dangerous behaviors and safe behaviors, observing and handling workers' behaviors at the production site of construction workers, enhancing the proactive response capabilities of construction workers, developing joint effects with construction workers on safety behaviors, correcting unsafe behaviors of construction workers, training safe behaviors, promoting the formation of a safety atmosphere, and improving safety performance management methods.

[0004] In the above traditional BBS-CI method, an observation action method based on behavioral safety is adopted, which can obtain a large amount of real first-hand information, but it takes a lot of manpower, material resources, financial resources and time, and the degree of intelligence is insufficient and the efficiency is low. Second, in the BBS-CI method, the intervention process (behavior correction) is based on the premise of key behaviors, and the analysis and determination of key behaviors are often obtained through empirical methods and behavioral observation methods. These two methods are highly subjective and cannot comprehensively and real-time collect and analyze key behaviors. There is a lack of in-depth analysis and summary of the characteristics and laws of unsafe behaviors of construction workers. Third, in the BBS-CI method, behavioral safety intervention mainly adopts three means: material rewards, goal setting and performance feedback, and management intervention. The "cast-net" and "one-size-fits-all" intervention methods have low accuracy and cannot fully capture the sites where unsafe behaviors are generated for precise intervention, and the effect is not good. Fourth, studies have shown that simple behavioral observation and performance feedback cannot continuously reduce safety accidents. They can only have a temporary impact and cannot fundamentally correct workers' unsafe behaviors. After the intervention, the root causes will begin to affect the unsafe behaviors again.

[0005] With the widespread application of artificial intelligence (AI) technology, behavioral safety management activities in the construction industry have received more powerful support. Artificial intelligence technology has a high degree of intelligence, automation and integration. Applying it in behavioral safety management activities in the construction industry can achieve efficient identification and classification of construction workers' behavioral safety, in-depth analysis and summary of unsafe behavior characteristics and patterns, so as to fully capture the sites where unsafe behaviors are generated for precise intervention, achieve targeted behavioral correction (behavioral training), dynamically match the effects of training and correction, and ultimately promote the formation of a safe atmosphere, improve safety performance and achieve overall safety goals. Summary of the invention

[0006] The purpose of the present invention is to provide a construction worker behavior safety management and correction method and system based on AI technology.

[0007] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is:

[0008] A construction worker behavior safety management and correction method and system based on AI technology, including an enterprise-level cloud computing center, multiple site-level regional edge computing centers and several front-end sensing devices;

[0009] The front-end sensing device is located at the front end or access end of the system. The front-end sensing device includes a fixed sensing device directly connected by a line and a mobile sensing device connected by a wireless signal. The video image acquisition module is used to obtain videos and pictures of unsafe behaviors of construction workers. The cloud computing center is located at the far end of the system, the regional edge computing center is located between the front-end sensing device and the cloud computing center, and the construction worker intelligent application terminal is located at the application end of the system. The regional edge computing center has the automatic recognition function, classification function and storage function of unsafe behaviors of construction workers.

[0010] The cloud computing center has AI intelligent learning function, unsafe behavior recognition model training function and behavior correction precision intervention function. The platform of the unsafe behavior recognition model training function imports the video image data set of unsafe behavior instances collected by the front-end sensing equipment from the construction site, repeatedly trains the unsafe behavior algorithm to obtain the unsafe behavior model, and then directly loads the model to the regional edge computing center, so as to perform artificial intelligence analysis on the video image data set, intelligently identify the trained unsafe behaviors, classify and store them; and perform sample annotation on the identified images, on the one hand, provide them to the artificial intelligence model training platform for training, and on the other hand, provide the required case samples to the platform of the behavior correction precision intervention function;

[0011] The platform with precise behavior correction intervention function uses data mining and statistical analysis in AI intelligent learning methods to identify the intervention nodes after the multi-dimensional interaction of unsafe behaviors with various job positions and construction processes, thereby realizing behavior correction or personalized training for construction workers to achieve the purpose of precise intervention. It also develops a personalized training system to correct workers' unsafe behaviors.

[0012] Preferably, the construction worker behavior safety management and correction method and system adopt a cloud-edge-end network system, which includes a front-end perception layer, a regional edge computing layer, a cloud computing layer, and an intelligent application layer;

[0013] The front-end perception layer is located at the front end or access end of the network system and is responsible for the network transmission of video image data collected at the construction site;

[0014] The regional edge computing layer is responsible for aggregating the unstructured video data and IoT data sent by each front-end sensing device, pre-processing it, triggering corresponding actions according to established rules, and uploading the processing structure and related data to the cloud;

[0015] The cloud computing layer is located at the far end of the system. It is responsible for processing and storing global information, taking on computing tasks that the regional edge computing layer cannot perform, and issuing business rules, algorithm models, and providing standard APIs for the development and docking of various applications to the regional edge computing layer. The cloud computing layer is also responsible for correcting workers' behavior, developing personalized behavioral safety training, and determining targeted intervention nodes based on the monitoring results of key behaviors of job types and construction processes, so as to achieve secondary and multiple interventions for unsafe behaviors.

[0016] The intelligent application layer uses the structured data that is analyzed and processed, combined with specific business needs and application models, to provide vertical application services for users.

[0017] Preferably, the cloud computing center is provided with a regional edge management module, a video cloud platform module, an unsafe behavior identification platform module, a behavior correction platform module and an Internet of Things platform module, and each module cooperates with each other to realize the AI ​​intelligent learning function, the unsafe behavior identification model training function and the behavior correction precise intervention function of the cloud computing center.

[0018] Preferably, the regional edge management module includes a regional management submodule, a component management submodule, a service management submodule, and a service orchestration submodule to implement resource monitoring functions of regional edge computing;

[0019] The video cloud platform module includes a video storage submodule, a video distribution submodule, a video processing submodule, and a quality diagnosis submodule to realize the structured data function of the video image;

[0020] The unsafe behavior identification platform module includes an algorithm market submodule, a model training submodule, a sample annotation submodule, and a model distribution submodule to realize the intelligent management function of the model algorithm;

[0021] The behavior correction platform module includes a user management submodule, a case management submodule, a feature analysis submodule, a content recommendation submodule, a correction feedback submodule, and a training evaluation submodule, which realizes the behavior correction of construction workers and achieves the purpose of precise intervention. A personalized training system is developed to ultimately correct the unsafe behavior of workers;

[0022] The Internet of Things platform module includes a device management submodule, a connection management submodule, a data analysis submodule, a data management submodule, a rule engine submodule, and a user management submodule to realize the management function of various types of devices.

[0023] Preferably, the regional edge computing center is provided with an edge computing network computing storage module, an operating system integrated platform module, a Docker container module, an application module, a device management module, and a scheduling management module;

[0024] The operating system integrated platform module includes a network submodule, a computing submodule, a storage submodule, and an operating system submodule to realize the edge computing support platform function;

[0025] The Docker container module is used to enable lightweight and modular use of multiple applications in a virtual machine;

[0026] The application module includes several unsafe behavior identification applications to realize the identification of various unsafe behaviors, and the device management module realizes the management function of access devices;

[0027] The scheduling management module is used to realize the task scheduling function, including the task scheduling of the cloud computing center and the regional edge computing center.

[0028] Preferably, the unsafe behavior identification platform module performs model training through an unsafe behavior model training method, and the model training method includes the following:

[0029] Step A1, first sort out the unsafe behaviors of construction workers during construction: The unsafe behaviors of construction workers during construction are divided into 13 categories with a total of 51 types. The following 51 unsafe behaviors are uniformly coded or numbered for classification of collected image data;

[0030] Step A2, video image acquisition and preprocessing: the video images taken by the fixed sensing equipment deployed at the construction site and the on-site photos taken by the inspector's mobile sensing equipment are preprocessed by removing noise, grayscale transformation, and geometric correction to reduce the impact of the construction environment of the construction site on the video analysis;

[0031] Step A3, case collection: according to the unsafe behaviors defined in step A1, unsafe behavior cases are collected in the preprocessed images;

[0032] Step A4, image sample annotation: manually annotate the target rectangular box of the unsafe behavior case, perform normalization processing and save relevant information of the target rectangular box, including the starting position, width and height;

[0033] Step A5, image sample storage: storing the labeled image samples to form a training image data set;

[0034] Step A6, model training: Model training includes several layers of convolutional layers and some downsampling and nonlinear transformations. During model training, the training video image data set is abstracted level by level, and finally a model feature with a certain discrimination ability for some images outside of training is obtained;

[0035] After a series of processing and convolutional network processing, the input video image retains valuable information and discards redundant information, that is, the image features are abstracted, and the interaction of a large amount of image feature information is used to train a model output;

[0036] Step A7, data post-processing: perform some post-processing and sorting on the results obtained by the convolutional network, and store them for use in the next step;

[0037] Step A8, feedback: adjust the weights obtained by the convolutional network, optimize the cost function using the gradient descent method, and adjust the weights of each layer of the convolutional network step by step from back to front, so that the generalization ability of the output model is greatly improved;

[0038] Step A9, model output: After continuous feedback, model training and data processing, the solution is optimized to obtain the final model result.

[0039] Preferably, in the system with AI intelligent learning function, the trained unsafe behaviors are intelligently identified, classified and stored, and a method for intelligently identifying, classifying and storing unsafe behaviors is adopted, and the steps of the method include the following:

[0040] Step B1, model distribution: the cloud computing center distributes the final model results obtained in step A9 to the regional edge computing center;

[0041] Step B2, model loading: The regional edge computing center loads the model results in B1. The loaded model results are the weight values ​​of each network layer obtained by the above-mentioned unsafe behavior model training method. The weight values ​​include millions of weight value parameters. The unsafe behavior recognition model of the entire recognition process is loaded to the regional edge computing center at the beginning of the system. The unsafe behavior recognition model is dynamically updated according to the passage of time and the rules of the model distribution submodule;

[0042] Step B3, behavior recognition: After loading is completed, a behavior recognition network layer is added to constrain the recognition results. The behavior recognition marks the location of unsafe behaviors and predicts the category of unsafe behaviors at the same time to confirm the credibility of detection and recognition; automatic behavior recognition and classification are performed on the video images from step A2;

[0043] Step B4, image annotation: After the behavior is identified, the output result is a certain unsafe behavior of the construction worker, and the identified network image is annotated;

[0044] Step B5, image sample storage: the recognized images and the annotated images are encoded and stored in the background in the regional edge computing center, the unsafe behavior recognition model is dynamically updated according to the rules of the model distribution submodule, and the updated image samples are deleted regularly;

[0045] Step B6, providing samples for model training: the image samples encoded and stored in the background of the regional edge computing center are transmitted to the cloud data center in real time to provide image samples for unsafe behavior model training, and can also be used as case samples required for behavior correction;

[0046] Step B7, data post-processing: The obtained image samples can be retrieved through standard retrieval tools, and can query specific events that have been recorded in real time; in addition, the image samples are also used as a big data source for the behavior correction platform to analyze unsafe behaviors, and realize the statistics and analysis of the unsafe behavior characteristics of construction workers;

[0047] Step B8: The updated model sample is output, and the output model is fed back into the unsafe behavior model training method.

[0048] Preferably, the method steps for statistics and analysis of unsafe behavior characteristics of construction workers in B7 include the following:

[0049] Step C1: Workers' unsafe behavior activities can collect a large amount of workers' unsafe behavior video image data from the on-site photos obtained in A2 as the basis for mining the statistical laws of construction workers' unsafe behavior;

[0050] Step C2, analyzing the unsafe behavior video image data in C1 by statistical analysis method, and mining the correlation between the unsafe behavior of construction workers and other dimensions; the dimensions include time dimension, space dimension, position dimension, and process dimension;

[0051] The time dimension indicates the time when the worker's unsafe behavior occurs; the space dimension indicates the spatial location where the worker's unsafe behavior occurs; the position dimension indicates the type of work position corresponding to the worker's unsafe behavior; the process dimension indicates the process stage where the worker's unsafe behavior occurs;

[0052] Step C3: Analyze the distribution characteristics of the collected data in different dimensions through a data mining algorithm to obtain multi-dimensional association rules of the collected data, obtain frequent item sets in the collected data through an iterative algorithm, and then construct multi-dimensional association rules of the collected data according to the confidence of the frequent item sets.

[0053] Preferably, the platform for the precise intervention function of behavior correction adopts a personalized behavior correction training method, and the method steps include the following:

[0054] Step D1, data collection: data collection of worker unsafe behavior videos, worker attribute data, worker safety level test scores, and behavior correction test scores;

[0055] Step D2, characteristic law analysis;

[0056] Step D3, behavior correction;

[0057] Step D4, behavior observation: using fixed sensing devices and mobile devices deployed at the construction site to automatically monitor the behavior of construction workers;

[0058] Step D5, data re-collection: Based on the automatic behavior monitoring method of D4, the video image data of unsafe behaviors of workers in different positions and construction processes are automatically collected in real time and continuously to achieve the acquisition and evaluation of the behavior correction of construction workers.

[0059] Preferably, the behavior correction in step D3 adopts a content-based recommendation training course, and the training course method includes the following:

[0060] First, feature information is extracted from the semantic information contained in the behavior correction content through a content analyzer, including the type of unsafe behavior, the type of work position corresponding to the unsafe behavior, the construction process, and the difficulty coefficient of generating behavior correction content based on the pictures reflecting the unsafe behavior of construction workers;

[0061] Then, characteristic information of construction workers is collected, including job positions, construction tasks, and information collected by construction workers during the behavior correction process. These feedback information are extracted through the file learner;

[0062] Finally, the corresponding relationship between the characteristic information of construction workers and the behavior correction content is established through the filtering component, and a set of the most relevant behavior correction content is recommended to the construction workers and pushed to the workers for learning.

[0063] The beneficial effects of the present invention are:

[0064] The system method in the present invention uses front-end sensing equipment (video sensing equipment) at the construction site, observation by safety inspectors (photographing and recording with smart phones), and training and evaluation, etc., to automatically observe and monitor key behaviors of job positions based on measurable, controllable, and observable standards; feedback on the results of behavioral observation and monitoring, timely exchange and communicate data information, analyze and mine the collected pan-scenario data, and further explore the key behavioral characteristics of different job positions, determine targeted intervention nodes, and implement secondary and multiple interventions for unsafe behaviors, ultimately achieving the purpose of preventing accidents and improving safety performance. This system method effectively reduces unsafe behaviors of construction workers during production operations and strengthens behavioral safety, thereby achieving the purpose of preventing accidents and improving safety performance. This method can not only solve the real-time, accuracy, and sustainability problems of the traditional BBS-CI method, but also meet the scientific and intelligent needs of modern buildings. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0066] Figure 1 It is a system module diagram of a construction worker behavior safety management correction method and system based on AI technology.

[0067] Figure 2 It is a flowchart diagram of the unsafe behavior model training method.

[0068] Figure 3 It is a flowchart diagram of a method for intelligently identifying, classifying and storing unsafe behaviors.

[0069] Figure 4 It is a flowchart of the program for statistics and analysis of unsafe behavior characteristics of construction workers.

[0070] Figure 5 It is a schematic diagram of the sequence numbers and time intervals of various unsafe behavior events of construction workers in subway construction in Example 5.

[0071] Figure 6 It is a three-dimensional line chart of unsafe behaviors of construction workers in subway construction in Example 5.

[0072] Figure 7 It is a platform with precise intervention function for behavior correction, and adopts personalized training methods for behavior correction.

[0073] Figure 8 This is a content-based personalized recommendation flowchart. DETAILED DESCRIPTION

[0074] The present invention provides a construction worker behavior safety management correction method and system based on AI technology. In order to make the purpose, technical solution and effect of the present invention clearer and more specific, the present invention is further described in detail below. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0075] The present invention is described in detail below in conjunction with the accompanying drawings:

[0076] Example 1

[0077] Combination Figures 1 to 8 , a construction worker behavior safety management and correction method and system based on AI technology, including an enterprise-level cloud computing center, multiple site-level regional edge computing centers and several front-end sensing devices;

[0078] The front-end sensing device is located at the front end or access end of the system. The front-end sensing device includes fixed sensing devices directly connected through lines and mobile sensing devices connected through wireless signals. The front-end sensing device is provided with a video image acquisition module. The video image acquisition module is used to obtain videos and pictures of unsafe behaviors of construction workers and perform video image acquisition in real time.

[0079] The cloud computing center is located at the far end of the system, the regional edge computing center is located between the front-end sensing device and the cloud computing center, and the construction worker intelligent application terminal is located at the application end of the system; the regional edge computing center has the automatic identification, classification and storage functions of unsafe behaviors of construction workers.

[0080] The cloud computing center has AI intelligent learning functions, unsafe behavior identification model training functions, and behavior correction precision intervention functions. The platform for unsafe behavior identification model training function imports video image data sets of unsafe behavior instances collected by front-end sensing devices from construction sites, repeatedly trains the unsafe behavior algorithm to obtain an unsafe behavior model, and then directly loads the model to the regional edge computing center, thereby performing artificial intelligence analysis on video images taken by video sensing equipment and safety inspectors' smartphones, intelligently identifying the trained unsafe behaviors, classifying them, and storing them; and annotating samples on the identified images, providing them to the artificial intelligence model training platform for training on the one hand, and providing them to the case samples required for behavior correction or personalized training on the other hand;

[0081] The platform with precise intervention function for behavior correction uses data mining and statistical analysis in AI intelligent deep learning methods to identify intervention nodes after multi-dimensional interactions between unsafe behaviors and various job positions and construction processes, thereby achieving behavior correction or personalized training for construction workers and achieving the purpose of precise intervention. A personalized training system has been developed to ultimately correct unsafe behaviors of workers.

[0082] The safety management and correction method of the present invention, based on the personalized training system, combines AI intelligent technology to establish a cyclic continuous process intervention system of data collection - characteristic law analysis - behavior correction - behavior observation - data re-collection; uses construction site video sensor equipment, smartphone photo recording and training evaluation during safety inspectors' observations, and automatically observes key behaviors of job positions based on measurable, controllable and observable standards; feeds back behavioral observation results, exchanges and communicates data information in a timely manner, analyzes and mines the collected pan-scenario data, explores the characteristic laws of key behaviors of different job positions again, determines targeted intervention nodes, and implements secondary and multiple interventions on unsafe behaviors, ultimately achieving the purpose of preventing accidents and improving safety performance.

[0083] The construction worker behavior safety management correction method and system adopts a cloud-edge network system, which includes a front-end perception layer, a regional edge computing layer, a cloud computing layer, and an intelligent application layer. The front-end perception layer is located at the front end or access end of the network system, and is responsible for the network transmission of video image data collected at the construction site; the regional edge computing layer is responsible for aggregating and pre-processing the unstructured video data and IoT data sent by each front-end perception device, triggering corresponding actions according to established rules, and uploading the processing structure and related data to the cloud.

[0084] The cloud computing layer is located at the far end of the entire network system; the cloud computing layer is responsible for the processing and storage of global information, undertakes computing tasks that the regional edge computing layer cannot perform, and sends business rules, algorithm models, and provides standard APIs for the development and docking of various applications to the regional edge computing layer. In addition, the cloud computing layer is also responsible for the behavior correction of workers, the development of personalized behavioral safety training, and the determination of targeted intervention nodes based on the monitoring results of key behaviors of job types and construction processes, so as to achieve secondary and multiple interventions for unsafe behaviors. The intelligent application layer uses the structured data that is analyzed and processed, combined with specific business needs and application models, to provide users with vertical application services.

[0085] The cloud computing center is equipped with regional edge management modules, video cloud platform modules, unsafe behavior identification platform modules, behavior correction platform modules, and Internet of Things platform modules. Each module works together to realize the AI ​​intelligent learning function, unsafe behavior identification model training function, and behavior correction precision intervention function of the cloud computing center.

[0086] The regional edge management module includes the regional management submodule, component management submodule, service management submodule, and service orchestration submodule to realize the resource monitoring function of regional edge computing. The video cloud platform module includes the video storage submodule, video distribution submodule, video processing submodule, and quality diagnosis submodule to realize the structured data function of video images. The unsafe behavior identification platform module includes the algorithm market submodule, model training submodule, sample annotation submodule, and model distribution submodule to realize the intelligent management function of the model algorithm;

[0087] The behavior correction platform module includes user management submodule, case management submodule, feature analysis submodule, content recommendation submodule, correction feedback submodule, and training evaluation submodule, which can realize the behavior correction of construction workers and achieve the purpose of precise intervention. A personalized training system has been developed to ultimately correct the unsafe behavior of workers. The IoT platform module includes device management submodule, connection management submodule, data analysis submodule, data management submodule, rule engine submodule, and user management submodule to realize the management function of various types of equipment.

[0088] The regional edge computing center is equipped with edge computing network computing storage module, operating system integrated platform module, Docker container module, application module, device management module, and scheduling management module.

[0089] The operating system integrated platform module includes a network sub-module, a computing sub-module, a storage sub-module, and an operating system sub-module to realize the edge computing support platform function.

[0090] The Docker container module enables the lightweight and modular use of multiple applications in virtual machines. The application module includes an unsafe behavior identification application to identify various unsafe behaviors, and the device management module to manage access devices. The scheduling management module implements task scheduling functions, including task scheduling of cloud computing centers and regional edge computing centers.

[0091] The unsafe behavior identification platform module performs model training through an unsafe behavior model training method, and its model training method includes the following:

[0092] Step A1: First, sort out the unsafe behaviors of construction workers during construction, and define i types of unsafe behaviors, 0<i≤I:

[0093] Unsafe behaviors of construction workers during construction. According to the Classification Standards for Casualty Accidents of Enterprise Employees, the unsafe behaviors of people in production activities are divided into 13 categories, totaling 51 types. The following 51 unsafe behaviors are uniformly coded or numbered for the classified image data;

[0094] Step A2, video image acquisition and preprocessing: the video images taken by the fixed sensing equipment deployed at the construction site and the on-site photos taken by the mobile sensing equipment of the full inspector are preprocessed by removing noise, grayscale transformation, and geometric correction to reduce the impact of the construction environment of the construction site on the video analysis;

[0095] Step A3, case collection: according to the i types of unsafe behaviors defined in step A1, unsafe behavior cases are collected in the preprocessed images;

[0096] Step A4, image sample annotation: manually annotate the target rectangular box of the unsafe behavior case, perform normalization processing and save relevant information of the target rectangular box, including the starting position, width and height;

[0097] Step A5, image sample storage: storing the labeled image samples to form a training image data set;

[0098] Step A6, model training: Model training includes several layers of convolutional layers and some downsampling and nonlinear transformations. During model training, the training video image data set is abstracted level by level, and finally a model feature with a certain discrimination ability for some images outside of training is obtained;

[0099] After a series of processing, the input video image represents a gradual reduction in information, which also means that after being processed by the convolutional network, valuable information is preserved, while some redundant information is discarded, that is, the image features are abstracted, and the interaction of a large amount of image feature information is used to jointly train a model output;

[0100] Step A7, data post-processing: perform some post-processing and sorting on the results obtained by the convolutional network, and store them for use in the next step;

[0101] Step A8, feedback: adjust the weights obtained by the convolutional network, optimize the cost function using the gradient descent method, and adjust the weights of each layer of the convolutional network step by step from back to front, so that the generalization ability of the output model is greatly improved;

[0102] Step A9, model output: After continuous feedback, model training and data processing, the solution is optimized to obtain the final model result.

[0103] In the system with AI intelligent learning function, the trained unsafe behaviors are intelligently identified, classified and stored. A method for intelligently identifying, classifying and storing unsafe behaviors is adopted, and the method steps include the following:

[0104] Step B1, model distribution: the cloud computing center distributes the final model results obtained in step A9 to the regional edge computing center;

[0105] Step B2, model loading: The regional edge computing center loads the model results in B1. The loaded model results are the weight values ​​of each network layer obtained by the above-mentioned unsafe behavior model training method. The weight values ​​include millions of weight value parameters. The unsafe behavior recognition model of the entire recognition process is loaded to the regional edge computing center at the beginning of the system. The unsafe behavior recognition model is dynamically updated according to the passage of time and the rules of the model distribution submodule;

[0106] Step B3, behavior recognition: After loading is completed, a behavior recognition network layer is added to constrain the recognition results. The behavior recognition marks the location of unsafe behaviors and predicts the category of unsafe behaviors at the same time to confirm the credibility of detection and recognition; automatic behavior recognition and classification are performed on the video images from the fixed sensing devices and mobile sensing devices in step A2;

[0107] Step B4, image annotation: After the behavior is identified, the output result is a certain unsafe behavior of the construction worker, and the identified network image is annotated;

[0108] Step B5, image sample storage: the recognized images and the annotated images are encoded and stored in the background in the regional edge computing center, the unsafe behavior recognition model is dynamically updated according to the rules of the model distribution submodule, and the updated image samples are deleted regularly;

[0109] Step B6, providing samples for model training: The image samples encoded and stored in the background of the regional edge computing center are transmitted to the cloud data center in real time to provide image samples for unsafe behavior model training, and can also be used as case samples required for behavior correction (personalized training);

[0110] Step B7, data post-processing: The obtained image samples can be retrieved through standard retrieval tools, and real-time query of specific events that have been recorded can be achieved; in addition, the image samples can also be used as a big data source for the behavior correction platform (personalized training platform) to analyze unsafe behaviors, and realize the statistics and analysis of the unsafe behavior characteristics of construction workers;

[0111] Step B8: The updated model sample is output, and the output model is fed back into the unsafe behavior model training method.

[0112] The steps of statistics and analysis of unsafe behavior characteristics of construction workers in B7 include the following:

[0113] Step C1: The worker unsafe behavior activity can collect a large amount of worker unsafe behavior video image data from the on-site photos obtained in A2 as a basis for mining the statistical laws of construction workers' unsafe behaviors;

[0114] Step C2, analyzing the unsafe behavior video image data in C1 by statistical analysis method, and mining the correlation between the unsafe behavior of construction workers and other dimensions; the dimensions include time dimension, space dimension, position dimension, and process dimension;

[0115] The time dimension indicates the time when the worker's unsafe behavior occurs; the space dimension indicates the spatial location where the worker's unsafe behavior occurs; the position dimension indicates the type of work position corresponding to the worker's unsafe behavior; the process dimension indicates the process stage where the worker's unsafe behavior occurs;

[0116] Step C3: Analyze the distribution characteristics of the collected data in different dimensions through a data mining algorithm to obtain multi-dimensional association rules of the collected data, obtain frequent item sets in the collected data through an iterative algorithm, and then construct multi-dimensional association rules of the collected data according to the confidence of the frequent item sets.

[0117] The platform for correcting the precise intervention function adopts a personalized training method for behavior correction, and the method steps include the following:

[0118] Step D1, data collection: data collection of worker unsafe behavior videos, worker attribute data, worker safety level test scores, and behavior correction test scores;

[0119] Step D2, characteristic law analysis;

[0120] Step D3, behavior correction: Behavior correction adopts a content-based recommendation training course, and the course method includes the following:

[0121] First, feature information is extracted from the semantic information contained in the behavior correction content through a content analyzer, including the type of unsafe behavior, the type of work position corresponding to the unsafe behavior, the construction process, and the difficulty coefficient of generating behavior correction content based on the pictures reflecting the unsafe behavior of construction workers;

[0122] Then, characteristic information of construction workers is collected, including job positions, construction tasks, and information collected by construction workers during the behavior correction process. These feedback information are extracted through the file learner;

[0123] Finally, the corresponding relationship between the characteristic information of construction workers and the behavior correction content is established through the filtering component, and a set of the most relevant behavior correction content is recommended to the construction workers and pushed to the workers for learning.

[0124] Step D4, behavior observation: Use fixed sensing equipment and mobile equipment deployed at the construction site to automatically monitor the behavior of construction workers. Step D5, data re-collection: Based on the behavior automatic monitoring method of D4, the video image data of unsafe behaviors of workers in different positions and construction processes are automatically collected in real time and continuously to achieve the acquisition and evaluation of the behavior correction of construction workers. The construction worker behavior safety management and correction method and system are also equipped with a worker information positioning system, which includes a worker safety helmet information positioning system. Workers in the area register their safety helmet information, and a GPS signal locator is set in the safety helmet, and a personal information QR code is set on the surface of the safety helmet.

[0125] Example 2

[0126] For a certain airport terminal and airport runway project under construction, it is necessary to monitor the safety behavior of the terminal construction site and runways on a large scale. The project adopts a construction worker behavior safety management and correction method and system based on AI technology. The construction worker behavior safety management and correction method and system based on AI technology includes an enterprise-level cloud computing center, multiple site-level regional edge computing centers, several front-end sensing devices and numerous construction worker intelligent application terminals.

[0127] The front-end sensing device is located at the front end or access end of the system, the cloud computing center is located at the remote end of the system, the regional edge computing center is located between the front-end sensing device and the cloud computing center, and the construction worker intelligent application terminal is located at the application end of the system.

[0128] There are two ways to connect the front-end sensing device to the regional edge computing center. The fixed sensing device (video sensor device) is connected directly via CAN / RJ45 / RS485 / RS232, and the mobile sensing device (smartphone) is connected via 5G / 4G mobile connection; the regional edge computing center and the cloud computing center are remotely connected via 5G / 4G / LAN / WLAN; the construction worker intelligent application terminal and the cloud computing center are connected via 5G / 4G mobile connection.

[0129] The front-end sensing device has a video image acquisition module, which includes several video sensor devices deployed at the construction site and several smartphones of safety inspectors to obtain videos and pictures of unsafe behaviors of construction workers, realizing real-time video image acquisition function.

[0130] The regional edge computing center has the functions of automatic identification and classification of unsafe behaviors of construction workers and storage. The cloud computing center has AI intelligent learning function, which can realize unsafe behavior identification model training and precise intervention (behavior correction) function.

[0131] The cloud computing center has the function of training unsafe behavior recognition models. The platform imports video image data sets of unsafe behavior instances collected by front-end sensing devices at construction sites, repeatedly trains the unsafe behavior algorithm to obtain an unsafe behavior model, and then directly loads the model to the regional edge computing center, thereby performing artificial intelligence analysis on the video images taken by the video sensing devices from the front-end sensing layer and the smartphones of the safety inspectors, intelligently identifying the trained unsafe behaviors, classifying them, and storing them. Samples are annotated on the identified images, which are provided to the artificial intelligence model training platform for training on the one hand, and to the case samples required for behavior correction (personalized training) on ​​the other.

[0132] The cloud computing center has the function of behavioral correction (personalized training) to achieve precise intervention. The platform uses AI intelligent deep learning methods such as data mining and statistical analysis to identify the intervention nodes after multi-dimensional interactions between unsafe behaviors and various job positions, construction processes (stages), etc., to achieve behavioral correction (personalized training) for construction workers and achieve the purpose of precise intervention. A personalized training system has been developed to ultimately correct the unsafe behaviors of workers.

[0133] On the basis of the "personalized training system", combined with AI intelligent technology, a cyclic continuous process intervention system of "data collection-characteristic law analysis-behavior correction (precise intervention)-behavior observation-data re-collection" is established. Adopting front-end sensing equipment (video sensing equipment) of construction sites, observation by safety inspectors (photographing and recording with smart phones) and training evaluation, the key behaviors of job positions are automatically observed (monitored) based on the standards of measurability, controllability and observability. Feedback the results of behavioral observation (monitoring), exchange and communicate data information in a timely manner, analyze and mine the collected pan-scenario data, explore the key behavioral characteristics and laws of different job positions again, determine the targeted intervention nodes, and implement secondary and multiple interventions for unsafe behaviors, so as to ultimately prevent accidents and improve safety performance.

[0134] The new generation of AI-based construction worker behavior safety management and correction system adopts a "cloud-edge-end" network system, which includes a front-end perception layer, a regional edge computing layer, a cloud computing layer, and an intelligent application layer. The front-end perception layer is located at the front end or access end of the network system, and is responsible for the collection of on-site video image data at the construction site; the access terminals of the front-end perception layer include various video perception devices, smart phones of safety inspectors, etc. The regional edge computing center is deployed close to the data source according to actual needs to provide the nearest service. The regional edge computing layer is located between the front-end perception device and the cloud computing center; the regional edge computing layer is responsible for aggregating and pre-processing the unstructured video data and IoT data sent by each front-end perception device, triggering corresponding actions according to established rules, and uploading the processing structure and related data to the cloud.

[0135] The cloud computing layer is located at the far end of the system; the cloud computing layer is responsible for processing and storing global information, undertakes computing tasks that the regional edge computing layer cannot perform, and sends business rules, algorithm models, and provides standard APIs for the development and docking of various applications to the regional edge computing layer.

[0136] In addition, the cloud computing layer is also responsible for correcting workers' behavior (precise intervention), developing personalized behavioral safety training, determining targeted intervention nodes based on the monitoring results of key behaviors in job positions and construction processes (stages), and achieving secondary and multiple interventions in unsafe behaviors, ultimately achieving the goal of preventing accidents and improving safety performance.

[0137] The intelligent application layer uses the structured data that is analyzed and processed, combined with specific business needs and application models, to provide users with vertical application services such as behavior recognition, safety management, and behavior correction (precision intervention). This layer of the present invention mainly provides application services such as intelligent recognition and classification of unsafe behaviors of construction workers, behavior correction (precision intervention), etc.

[0138] The cloud computing center is equipped with a regional edge management module, a video cloud platform module, an unsafe behavior identification platform module, a behavior correction platform (personalized training platform) module, and an Internet of Things platform module. The regional edge management module includes a regional management submodule, a component management submodule, a service management submodule, and a service orchestration submodule to realize the resource monitoring function of regional edge computing. The video cloud platform module includes a video storage submodule, a video distribution submodule, a video processing submodule, and a quality diagnosis submodule to realize the structured data function of video images; the unsafe behavior identification platform module includes an algorithm market submodule, a model training submodule, a sample annotation submodule, and a model distribution submodule to realize the intelligent management function of the model algorithm.

[0139] The behavior correction platform (personalized training platform) module includes user management submodule, case management submodule, feature analysis submodule, content recommendation submodule, correction feedback submodule, and training evaluation submodule, which realizes the behavior correction (personalized training) of construction workers and achieves the purpose of precise intervention. A personalized training system is developed to ultimately correct the unsafe behavior of workers. The IoT platform module includes device management submodule, connection management submodule, data analysis submodule, data management submodule, rule engine submodule, and user management submodule to realize the management function of various types of equipment.

[0140] The regional edge computing center is equipped with an edge computing network computing storage module, an operating system integrated platform module, a Docker container module, an application module, a device management module, and a scheduling management module. The operating system integrated platform module includes a network submodule, a computing submodule, a storage submodule, and an operating system submodule to realize the edge computing support platform function. The Docker container module is used to realize the lightweight and modular use of multiple applications in a virtual machine; the application module includes i unsafe behavior identification applications to realize the identification of various unsafe behaviors. The device management module is used to realize the management function of the access device; the scheduling management module is used to realize the task scheduling function, including the task scheduling of the cloud computing center and the regional edge computing center.

[0141] Example 3

[0142] The unsafe behavior identification platform module in Example 2 is trained by an unsafe behavior model training method, and the model training method includes the following steps (see the attached Figure 3 shown):

[0143] First, the unsafe behaviors of construction workers during construction are sorted out, and i unsafe behaviors Actioni are defined, 0<i≤I.

[0144] Step 101, the definition and concept of unsafe behavior of construction workers are abstract, but the unsafe behavior of construction workers at the construction site is specific, observable and measurable. This type of unsafe behavior, construction site according to the "Enterprise Employee Casualty Accident Classification Standard" (GB6441-86) standard, can be divided into 13 categories of unsafe behaviors of people in production activities, a total of 51 types, see the following table (Table 1: "Enterprise Employee Casualty Accident Classification Standard" (GB6441-86) divides unsafe behaviors of people into 13 categories, a total of 51 types), the present invention is for the following 51 types of unsafe behavior list for unified coding (numbering), for classification of collected image data.

[0145] The following is a list of unsafe behaviors:

[0146]

[0147]

[0148]

[0149] Step 102, video image acquisition and preprocessing: the video images captured by the front-end sensing equipment video sensor equipment deployed at the construction site and the on-site photos taken by the smart phones of all inspectors are preprocessed by removing noise, grayscale conversion, and geometric correction to reduce the impact of the construction environment of the construction site on video analysis;

[0150] Taking the unsafe behavior activities at a construction site of an airport runway project as an example, the process of collecting image data for automatically monitoring the unsafe behavior activities of construction workers is described. The front-end sensing equipment deployed at the construction site is used. The video sensing equipment generally includes network cameras, POE power supply modules and switches, which constitute the hardware environment of the system. The network camera behavior analysis technology constitutes the software environment of the system. The video sensing equipment automatically monitors the construction site and obtains dynamic scene data.

[0151] The picture of unsafe behavior activities of construction workers at the construction site collected by smartphone is a screenshot of the unsafe behavior activities of construction workers collected by the smartphone of the safety inspector in a certain airport runway project. It is the interface of the smartphone APP, reflecting the collection process of such unsafe behavior photos at the airport construction. The safety inspector finds a scene reflecting the unsafe behavior at the construction site and takes a photo through the mobile phone APP. He describes the photo through text or voice according to the example requirements and submits the information after completion. The observer can also choose to continue collecting or exit.

[0152] Step 103, case collection: according to the i-type unsafe behavior Actioni defined in step A1, unsafe behavior cases are collected in the preprocessed image.

[0153] Step 104, image sample annotation: manually annotate the target rectangular box of the unsafe behavior case, perform normalization processing and save relevant information of the target rectangular box, the relevant information including the starting position, width and height.

[0154] Step 105, image sample storage: store the labeled image samples to form a training image data set Datai, 0<i≤I;

[0155] Step 106, model training; model training is the core of the unsafe behavior recognition model output. This module includes several layers of convolutional layers and some downsampling and nonlinear transformations; the training video image data set Datai is abstracted level by level, and finally a model feature (stored in the form of model weights) that has a certain discrimination ability for some images outside of training is obtained. After a series of processing, the input video image represents that the information is gradually reduced, which also means that after being processed by the convolutional network, valuable information is preserved, and some redundant information is discarded, that is, the image features are abstracted, and a large amount of image feature information interacts with each other to jointly train a model output Modeli, 0<i≤I.

[0156] Step 107, data post-processing: perform some post-processing and sorting on the results obtained by the convolutional network, and store them for the next step. Step 108, feedback (weight adjustment): adjust the weights obtained by the convolutional network, use the gradient descent method to optimize the cost function, and adjust the weights of each layer of the convolutional network step by step from back to front, so that the generalization ability of the output model is greatly improved; Step 109, model output: after continuous feedback, model training and data processing, optimize the scheme and obtain the final model result Modeli,0<i≤I.

[0157] Example 4

[0158] In Example 3, unsafe behaviors are intelligently identified using AI, and an intelligent identification, classification and storage method for unsafe behaviors is used during the identification process. The method includes the following steps:

[0159] Model distribution, the cloud computing center distributes the final model result Modeli in step 109 to the regional edge computing center;

[0160] Model loading: The regional edge computing center loads Modeli. The loaded model Modeli is the weight value of each network layer trained by the above-mentioned unsafe behavior model training method. The weight value includes millions of weight value parameters. The unsafe behavior recognition model of the entire recognition process is loaded to the regional edge computing center at the beginning of the system. The unsafe behavior recognition model is dynamically updated according to the passage of time and the rules of the model distribution sub-module.

[0161] Behavior recognition: After loading is complete, a behavior recognition network layer is added to constrain the recognition results. The behavior recognition marks the location of unsafe behaviors and predicts the category of unsafe behaviors at the same time to confirm the credibility of detection and recognition; automatic behavior recognition and classification are performed on the video images taken by the front-end perception device video sensor device and the security inspector's smartphone in step 102.

[0162] Image annotation: After behavior recognition, the output result is a certain unsafe behavior of construction workers, and the identified network image is annotated;

[0163] Image sample storage: The recognized and annotated images are encoded and stored in the background in the regional edge computing center. The unsafe behavior recognition model is dynamically updated according to the rules of the model distribution submodule, and the updated image samples are deleted regularly.

[0164] Providing samples for model training: Image samples encoded and stored in the background of the regional edge computing center are transmitted to the cloud data center in real time to provide image samples for unsafe behavior model training, and can also be used as case samples required for behavior correction (personalized training).

[0165] Data post-processing: The obtained image samples can be retrieved through standard retrieval tools, and can query specific events that have been recorded in real time; in addition, image samples can also be used as a big data source for behavior correction platforms (personalized training platforms) to analyze unsafe behaviors, and realize the statistics and analysis of unsafe behavior characteristics of construction workers.

[0166] The updated model samples are output, and the output model is fed back into the unsafe behavior model training method.

[0167] Example 5

[0168] The content and method of statistics and analysis of unsafe behavior characteristics of construction workers in Example 4 include the following:

[0169] Workers' unsafe behavior activities can collect a large amount of video image data of workers' unsafe behavior from videos taken by video sensing devices and on-site photos taken by smart phones, which serves as the basis for mining the statistical laws of construction workers' unsafe behavior. In the present invention, the corresponding steps are "video image acquisition and preprocessing" of A2 and "data post-processing" of B7, which will not be repeated here.

[0170] By analyzing the video image data of unsafe behaviors through statistical analysis, the correlation between the unsafe behaviors of construction workers and other dimensions can be explored. The dimensions include many, generally: time dimension, space dimension, position dimension, process dimension (construction stage dimension), etc. The time dimension indicates the time when the unsafe behavior of workers occurs; the space dimension indicates the spatial location where the unsafe behavior of workers occurs; the position dimension indicates the type of work corresponding to the unsafe behavior of workers; the process dimension indicates the process stage where the unsafe behavior of workers occurs. The collected data is analyzed by statistical analysis to obtain the distribution characteristics of the collected data in different dimensions.

[0171] (1) Correspondence between unsafe behavior data and worker job position data

[0172] The eight major types of work in the traditional construction industry include carpentry, reinforcement workers, plastering workers, masonry workers, concrete workers, scaffolding workers, painters and waterproofing workers. In combination with a specific project, such as subway construction, subway construction workers can be divided into general workers, scaffolding workers, reinforcement workers, formwork workers, welders, electricians, machine operators, etc. Among them, general workers are non-professional workers, and machine operators include crane drivers, excavator drivers, bulldozer drivers, etc.

[0173] Establish a corresponding relationship between unsafe behaviors and workers' job positions. Certain types of unsafe behaviors, such as scaffolding and formwork engineering, correspond to specific professional jobs such as scaffolders and formwork workers. Some types of unsafe behaviors correspond to all job positions and may occur to any worker at the subway construction site, such as civilized construction unsafe behaviors such as "workers entering the construction site without wearing safety helmets".

[0174] The corresponding relationship between some unsafe behaviors and the types of work of subway construction workers is shown in the following table (as shown in the following table: The corresponding relationship between some unsafe behaviors and the types of work of subway construction workers):

[0175] Serial number Unsafe behavior of workers Corresponding worker's job position 1 Workers smoking casually at the construction site all 2 Workers are not satisfied with the regulations on setting up scissor braces Scaffolder 3 There are still some hanging templates left Formwork 4 Unauthorized personnel entering the hazardous area during hoisting Machinery operators 5 The excavator is in an unstable position Machinery operators

[0176] (2) Correspondence between unsafe behavior data and time data

[0177] In combination with a specific project, taking the above-mentioned subway construction as an example, 9 types of unsafe behaviors are analyzed, including civilized construction, safety protection, track area construction safety, foundation pit support, scaffolding, mining tunnel construction, formwork engineering, lifting and hoisting, and shield tunnel construction. The event sequence-interval time (days) is plotted according to the order in which the workers' unsafe behaviors occur as the horizontal axis-the event sequence, and the time interval between the workers' unsafe behaviors as the vertical axis-the interval time (days), as shown in the attached figure. Figure 5 .

[0178] (3) Correspondence between unsafe behavior data and spatial data and construction process (stage) data

[0179] In "Image Sample Storage", the identified and annotated images are encoded, and a unified encoding rule can be used in combination with the construction process, such as "unit project-sub-unit project-division project-sub-item project-type-unsafe behavior" for recording, so the corresponding construction stage information when each unsafe behavior occurs can be obtained from the image data encoding of the cloud platform. Spatial data can be obtained based on the coverage area of ​​the video perception equipment deployed at the construction site and the location of the on-site photos taken by the safety inspectors.

[0180] C3. Analyze the distribution characteristics of the collected data in different dimensions through a data mining algorithm to obtain multi-dimensional association rules of the collected data. Obtain frequent item sets in the collected data through an iterative algorithm, and then construct multi-dimensional association rules of the collected data based on the confidence of the frequent item sets.

[0181] In combination with subway construction, the specific approach is as follows: to build an association rule mining database, we must first find out two corresponding relationships: one is the corresponding relationship between the unsafe behavior data reflected by the image and the workers' job position data, and the other is the corresponding relationship between the unsafe behavior data reflected by the image and the on-site construction stage data.

[0182] The data used for association rule mining were statistically analyzed, and the significance of the influence of job positions and construction stages on workers' unsafe behaviors was explained by non-repeated two-way variance analysis. According to the actual situation of a subway station construction, the job positions are divided into 7 categories: machine operators, electricians, welders, formwork workers, reinforcement workers, scaffold workers and general workers. The main construction stages corresponding to the sub-projects are divided into 8 categories: retaining structure, excavation and dewatering, support erection, bottom plate excavation, bottom plate pouring, middle plate pouring, top plate pouring and support removal. According to the above method, the correspondence between the unsafe behaviors of workers and job positions in the collected data on unsafe behaviors of subway station construction workers, as well as the correspondence between the time when the unsafe behaviors of workers occurred and the on-site construction stage were found. The recording results were plotted into a three-dimensional line graph with the worker's job position, construction stage and number of unsafe behaviors as the axis. It can be seen intuitively from the figure that the fluctuation of unsafe behaviors of machine operators in each construction stage is the largest, and the fluctuation of unsafe behaviors of general workers in each construction stage is the smallest, as shown in the attached figure. Figure 6 .

[0183] Example 6

[0184] The behavior correction platform uses AI intelligent deep learning methods such as data mining and statistical analysis to identify the intervention nodes after the interaction between unsafe behaviors of construction workers and various job types, construction processes, and other dimensions. It has developed a personalized training system and established a "data collection-characteristic law analysis-behavior correction-behavior observation-data re-collection" cycle continuous process intervention system to achieve precise intervention in unsafe behaviors of construction workers. Behavior correction (personalized training) includes the following steps:

[0185] Data collection. In addition to the collection of video image data of workers’ unsafe behaviors, it also includes worker attribute data, and feedback on workers’ safety level test scores and behavior correction test scores.

[0186] (1) The video image data collection of workers’ unsafe behaviors has been explained in the video image collection and preprocessing and data post-processing, and will not be repeated here.

[0187] (2) Worker attribute data collection. Individual factors such as the psychology and physiology of construction workers are important causes of unsafe behaviors of workers. On the other hand, the improvement of the assembly and mechanization level of construction projects and the requirements of smart buildings also require workers to have higher qualities to match them. Therefore, construction workers should establish and follow the standards for worker access management and select workers who meet the requirements to work in construction-related positions. Strengthen the access management of workers from the following aspects: a. Physiological requirements, including gender, age, physical characteristics, and occupational taboos; b. Psychological requirements, including work attitude, sense of responsibility, willpower, tolerance, and team spirit; c. Educational experience, including academic qualifications, diplomas, and work training experience; d. Work experience, including professional work experience and related work experience.

[0188] (3) Data collection of workers’ safety level test scores and behavior correction test score feedback is done, on the one hand, automatically through the worker client of the developed behavior safety correction system (personalized behavior safety training system). On the other hand, statistical data after behavior correction is obtained through the collection of video image data of workers’ unsafe behaviors.

[0189] Characteristic regularity analysis: The characteristic statistics and regularity analysis of unsafe behaviors of construction workers are described in detail in Example 5, which will not be repeated here.

[0190] Behavior correction. The present invention intends to adopt a content-based recommendation training course, such as the attached Figure 8 As shown in (Content-based personalized recommendation flow chart). First, through the content analyzer, feature information is extracted from the semantic information contained in the behavior correction content, including the type of unsafe behavior, the type of work position corresponding to the unsafe behavior, the construction process (stage), etc., as well as the difficulty coefficient of generating behavior correction content from the pictures of unsafe behavior of construction workers; then, the feature information of construction workers is collected, including the type of work position, construction tasks, and the information collected by construction workers during the behavior correction process (safety level test scores and behavior correction test scores). These feedback information are extracted through the file learner; finally, the corresponding relationship between the feature information of construction workers and the behavior correction content is established through the filtering component, and a set of behavior correction content with the greatest relevance is recommended to the construction workers, and pushed to the workers for learning.

[0191] Behavioral observation: The behavior of construction workers in training is observed by using front-end sensing devices, video sensing equipment deployed at the construction site, and the smartphones of all inspectors to automatically monitor (observe) the behavior of construction workers.

[0192] Data re-collection. Based on the means of automatic behavior monitoring (observation), the video image data of unsafe behaviors of workers in different positions and construction processes (stages) are automatically collected in real time and continuously to achieve the acquisition and evaluation of the behavior correction of construction workers. A continuous process intervention system of "data collection-characteristic law analysis-behavior correction-behavior observation-data re-collection" is established to achieve precise intervention in the unsafe behaviors of construction workers.

[0193] Example 7

[0194] The construction worker behavior safety management and correction method and system are also provided with a worker information positioning system, which includes a worker safety helmet information positioning system. Workers in the area register their safety helmet information, and a GPS signal locator is set in the safety helmet, and a personal information QR code is set on the surface of the safety helmet.

[0195] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A construction worker behavior safety management and correction system based on AI technology, characterized in that: It includes an enterprise-level cloud computing center, multiple construction site-level regional edge computing centers, and several front-end sensing devices; The front-end sensing device is located at the front end or access end of the system. The front-end sensing device includes a fixed sensing device directly connected by a line and a mobile sensing device connected by a wireless signal. The video image acquisition module is used to obtain videos and pictures of unsafe behaviors of construction workers. The cloud computing center is located at the far end of the system, the regional edge computing center is located between the front-end sensing device and the cloud computing center, and the construction worker intelligent application terminal is located at the application end of the system. The regional edge computing center has the functions of automatic identification, classification and storage of unsafe behaviors of construction workers; The cloud computing center has AI intelligent learning function, unsafe behavior recognition model training function and behavior correction precision intervention function. The platform of the unsafe behavior recognition model training function imports the video image data set of unsafe behavior instances collected by the front-end sensing equipment from the construction site, repeatedly trains the unsafe behavior algorithm to obtain the unsafe behavior model, and then directly loads the model to the regional edge computing center, so as to perform artificial intelligence analysis on the video image data set, intelligently identify the trained unsafe behaviors, classify and store them; and perform sample annotation on the identified images, on the one hand, provide them to the artificial intelligence model training platform for training, and on the other hand, provide the required case samples to the platform of the behavior correction precision intervention function; The platform with precise behavior correction and intervention functions uses data mining and statistical analysis in AI intelligent learning methods to identify intervention nodes after multi-dimensional interactions between unsafe behaviors and various job positions and construction processes, thereby achieving behavior correction or personalized training for construction workers and achieving the purpose of precise intervention. It also develops a personalized training system to correct workers' unsafe behaviors. The construction worker behavior safety management and correction system adopts a cloud-edge network system, which includes a front-end perception layer, a regional edge computing layer, a cloud computing layer, and an intelligent application layer; The front-end perception layer is located at the front end or access end of the network system and is responsible for the network transmission of video image data collected at the construction site; The regional edge computing layer is responsible for aggregating the unstructured video data and IoT data sent by each front-end sensing device, pre-processing it, triggering corresponding actions according to established rules, and uploading the processing structure and related data to the cloud; The cloud computing layer is located at the far end of the system. It is responsible for processing and storing global information, taking on computing tasks that the regional edge computing layer cannot perform, and issuing business rules, algorithm models, and providing standard APIs for the development and docking of various applications to the regional edge computing layer. The cloud computing layer is also responsible for correcting workers' behavior, developing personalized behavioral safety training, and determining targeted intervention nodes based on the monitoring results of key behaviors of job types and construction processes, so as to achieve secondary and multiple interventions for unsafe behaviors. The intelligent application layer uses the structured data processed by analysis, combined with specific business needs and application models, to provide users with vertical application services; The unsafe behavior identification platform module performs model training through an unsafe behavior model training method, and the model training method includes the following: Step A1, first sort out the unsafe behaviors of construction workers during construction: The unsafe behaviors of construction workers during construction are divided into 13 categories with a total of 51 types. The following 51 unsafe behaviors are uniformly coded or numbered for classification of collected image data; Step A2, video image acquisition and preprocessing: the video images taken by the fixed sensing equipment deployed at the construction site and the on-site photos taken by the inspector's mobile sensing equipment are preprocessed by removing noise, grayscale transformation, and geometric correction to reduce the impact of the construction environment of the construction site on the video analysis; Step A3, case collection: according to the unsafe behaviors defined in step A1, unsafe behavior cases are collected in the preprocessed images; Step A4, image sample annotation: manually annotate the target rectangular box of the unsafe behavior case, perform normalization processing and save relevant information of the target rectangular box, including the starting position, width and height; Step A5, image sample storage: storing the labeled image samples to form a training image data set; Step A6, model training: Model training includes several layers of convolutional layers and some downsampling and nonlinear transformations. During model training, the training video image data set is abstracted level by level, and finally a model feature with a certain discrimination ability for some images outside of training is obtained; After a series of processing and convolutional network processing, the input video image retains valuable information and discards redundant information, that is, the image features are abstracted, and the interaction of a large amount of image feature information is used to train a model output; Step A7, data post-processing: perform some post-processing and sorting on the results obtained by the convolutional network, and store them for use in the next step; Step A8, feedback: adjust the weights obtained by the convolutional network, optimize the cost function using the gradient descent method, and adjust the weights of each layer of the convolutional network step by step from back to front; Step A9, model output: After continuous feedback, model training and data processing, optimize the solution and obtain the final model result; In the system with AI intelligent learning function, the trained unsafe behaviors are intelligently identified, classified and stored, and a method for intelligently identifying, classifying and storing unsafe behaviors is adopted, and the steps of the method include the following: Step B1, model distribution: the cloud computing center distributes the final model results obtained in step A9 to the regional edge computing center; Step B2, model loading: The regional edge computing center loads the model results in B1. The loaded model results are the weight values ​​of each network layer obtained by the above-mentioned unsafe behavior model training method. The weight values ​​include millions of weight value parameters. The unsafe behavior recognition model of the entire recognition process is loaded to the regional edge computing center at the beginning of the system. The unsafe behavior recognition model is dynamically updated according to the passage of time and the rules of the model distribution submodule; Step B3, behavior recognition: After loading is completed, a behavior recognition network layer is added to constrain the recognition results. The behavior recognition marks the location of unsafe behaviors and predicts the category of unsafe behaviors at the same time to confirm the credibility of detection and recognition; automatic behavior recognition and classification are performed on the video images from step A2; Step B4, image annotation: After the behavior is identified, the output result is a certain unsafe behavior of the construction worker, and the identified network image is annotated; Step B5, image sample storage: the recognized images and the annotated images are encoded and stored in the background in the regional edge computing center, the unsafe behavior recognition model is dynamically updated according to the rules of the model distribution submodule, and the updated image samples are deleted regularly; Step B6, providing samples for model training: the image samples encoded and stored in the background of the regional edge computing center are transmitted to the cloud data center in real time to provide image samples for unsafe behavior model training as case samples required for behavior correction; Step B7, data post-processing: The obtained image samples can be retrieved through standard retrieval tools, and can query specific events that have been recorded in real time; in addition, the image samples are also used as a big data source for the behavior correction platform to analyze unsafe behaviors, and realize the statistics and analysis of the unsafe behavior characteristics of construction workers; Step B8: The updated model sample is output, and the output model is fed back into the unsafe behavior model training method.

2. According to claim 1, a construction worker behavior safety management and correction system based on AI technology is characterized in that: The cloud computing center is provided with a regional edge management module, a video cloud platform module, an unsafe behavior identification platform module, a behavior correction platform module and an Internet of Things platform module. Each module works together to realize the AI ​​intelligent learning function, the unsafe behavior identification model training function and the behavior correction precise intervention function of the cloud computing center.

3. According to claim 2, a construction worker behavior safety management and correction system based on AI technology is characterized in that: The regional edge management module includes a regional management submodule, a component management submodule, a service management submodule, and a service orchestration submodule to implement the resource monitoring function of regional edge computing; The video cloud platform module includes a video storage submodule, a video distribution submodule, a video processing submodule, and a quality diagnosis submodule to realize the structured data function of the video image; The unsafe behavior identification platform module includes an algorithm market submodule, a model training submodule, a sample annotation submodule, and a model distribution submodule to realize the intelligent management function of the model algorithm; The behavior correction platform module includes a user management submodule, a case management submodule, a feature analysis submodule, a content recommendation submodule, a correction feedback submodule, and a training evaluation submodule, which realizes the behavior correction of construction workers and achieves the purpose of precise intervention. A personalized training system is developed to ultimately correct the unsafe behavior of workers; The Internet of Things platform module includes a device management submodule, a connection management submodule, a data analysis submodule, a data management submodule, a rule engine submodule, and a user management submodule to realize the management function of various types of devices.

4. According to claim 1, a construction worker behavior safety management and correction system based on AI technology is characterized in that: The regional edge computing center is provided with an edge computing network computing storage module, an operating system integrated platform module, a Docker container module, an application module, a device management module, and a scheduling management module; The operating system integrated platform module includes a network submodule, a computing submodule, a storage submodule, and an operating system submodule to realize the edge computing support platform function; The Docker container module is used to enable lightweight and modular use of multiple applications in a virtual machine; The application module includes several unsafe behavior identification applications to realize the identification of various unsafe behaviors, and the device management module to realize the management function of access devices; The scheduling management module is used to realize the task scheduling function, including the task scheduling of the cloud computing center and the regional edge computing center.

5. According to claim 1, a construction worker behavior safety management and correction system based on AI technology is characterized in that: The method steps for statistics and analysis of unsafe behavior characteristics of construction workers in B7 include the following: Step C1: Workers' unsafe behavior activities. From the on-site photos obtained from A2, a large amount of video image data of workers' unsafe behavior is collected as the basis for mining the statistical laws of construction workers' unsafe behavior; Step C2, analyzing the unsafe behavior video image data in C1 by statistical analysis method, and mining the correlation between the unsafe behavior of construction workers and other dimensions; the dimensions include time dimension, space dimension, position dimension, and process dimension; The time dimension indicates the time when the worker's unsafe behavior occurs; the space dimension indicates the spatial location where the worker's unsafe behavior occurs; the position dimension indicates the type of work position corresponding to the worker's unsafe behavior; the process dimension indicates the process stage where the worker's unsafe behavior occurs; Step C3: Analyze the distribution characteristics of the collected data of different dimensions through a data mining algorithm to obtain multi-dimensional association rules of the collected data, obtain frequent item sets in the collected data through an iterative algorithm, and then construct multi-dimensional association rules of the collected data according to the confidence of the frequent item sets.

6. The construction worker behavior safety management and correction system based on AI technology according to claim 1 is characterized in that: The platform for precise behavior correction intervention function adopts a personalized behavior correction training method, and the method steps include the following: Step D1, data collection: data collection of worker unsafe behavior videos, worker attribute data, worker safety level test scores, and behavior correction test scores; Step D2, characteristic law analysis; Step D3, behavior correction; Step D4, behavior observation: using fixed sensing devices and mobile devices deployed at the construction site to automatically monitor the behavior of construction workers; Step D5, data re-collection: Based on the automatic behavior monitoring method of D4, the video image data of unsafe behaviors of workers in different positions and construction processes are automatically collected in real time and continuously to achieve the acquisition and evaluation of the behavior correction of construction workers.

7. The AI-based construction worker behavior safety management and correction system according to claim 6 is characterized in that: The behavior correction in step D3 adopts a content-based recommendation training course, and the training course method includes the following: First, feature information is extracted from the semantic information contained in the behavior correction content through a content analyzer, including the type of unsafe behavior, the type of work position corresponding to the unsafe behavior, the construction process, and the difficulty coefficient of generating behavior correction content based on the pictures reflecting the unsafe behavior of construction workers; Then, characteristic information of construction workers is collected, including job positions, construction tasks, and information collected by construction workers during the behavior correction process. These feedback information are extracted through the file learner; Finally, the corresponding relationship between the characteristic information of construction workers and the behavior correction content is established through the filtering component, and a set of the most relevant behavior correction content is recommended to the construction workers and pushed to the workers for learning.

Images