CN114782988A - Construction environment-oriented multi-stage safety early warning method - Google Patents

Abstract

The invention discloses a construction environment-oriented multi-stage safety early warning method, and belongs to the field of safety early warning. A multilevel safety early warning method facing to a construction environment adopts a hierarchical early warning mechanism, and improves the systematicness, accuracy and reliability of the multilevel safety early warning of construction big data; the acquired data are labeled and subjected to danger division, and the early warning cost is effectively saved on the premise of ensuring the accuracy and the timeliness by deeply mining the incidence relation among the multi-dimensional multi-element data; in the detection process, a visual depth analysis scheme suitable for project construction scenes is provided by combining the prior advanced algorithm, the invention effectively prevents the occurrence of dead angles in safety detection and endows the system with the capability of dynamically adjusting the early warning level.

Description

A multi-level safety early warning method for construction environment

technical field

The invention belongs to the field of safety early warning, in particular to a construction environment-oriented multi-level safety early warning method.

Background technique

With the continuous advancement of industrial information technology, how to carry out early warning and prevention and control for safety in the process of project construction in multiple directions, angles and levels has become one of the problems to be solved urgently in safety production. The data of the early warning mechanism mainly comes from the big data of the construction site, which focuses on the safety management and control of the construction site, and focuses on key elements such as personnel, machinery, materials, and the environment. However, the existing related early warning models have the disadvantages of single data, insufficient data linkage, Various problems such as small data processing scale, poor technical scalability, unclear hierarchical control boundaries, and high cost result in insufficient system migration and scalability, low accuracy and reliability, and lack of timeliness and linkage. Therefore, starting from specific construction scenarios, it is particularly important to design a general, reliable, systematic, and highly scalable early warning method.

Different from the general early warning system, the reasons for the ineffectiveness and poor usability of the project construction early warning mechanism are mainly divided into the following three aspects: First, the construction scene is complex and changeable, relying only on the data collected by sensors, resulting in a single source, and Due to its own limitations, there are loopholes in detection, the acquisition rate of effective information is low, and the data redundancy is large, and it is impossible to accurately classify, extract and determine the danger level of early warning information; secondly, people and the environment, objects (materials and equipment, etc.) The relationship with the environment and between people and things (materials and equipment, etc.) is ever-changing, and the construction scenes and states are different. It increases the difficulty of the early warning system; finally, in the construction scene, it is difficult to detect abnormal behaviors such as people fighting, smoking, and chatting in a conventional way. The existing early warning system lacks global monitoring and large-scale image capture, processing, and analysis. ability, which seriously affects the security early warning effect of construction big data.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the above-mentioned shortcomings of the prior art, and to provide a multi-level safety early warning method oriented to the construction environment.

To achieve the above object, the present invention adopts the following technical solutions to realize:

A multi-level safety early warning method oriented to construction environment, comprising the following steps;

S1. Collect multi-dimensional construction big data C, the multi-dimensional construction big data C includes image video and text data C _text of ambient temperature and humidity, and the image video is composed of historical data C _old to be marked and new data collected in real time C _new composition;

The historical data C _old to be labeled is labeled in different dimensions to obtain a labeled data set _CL ; the labeled data set _CL includes a dangerous area data set C _Da and a dangerous behavior data set C _Db ;

S2. Based on the label data set C _L , use the neural network to predict the wearing condition of the safety equipment, and obtain the wearing prediction result;

The dangerous area D _area is divided based on the dangerous area data set C _Da , and the edge detection confidence C _border is obtained by calculation; the personnel position information is predicted by using the neural network, and the personnel position detection confidence C _location is obtained;

The historical moving image MHI is generated based on the _new data Cnew collected in real time, the spatiotemporal information of normal human movement is obtained based on the dangerous behavior data set C _Db , and the abnormal behavior detection model is constructed and updated by using the neural network based on the spatiotemporal information and the historical moving image MHI. Predict human behavior;

Calculate the mean and standard deviation of the data according to the text data C _text of ambient temperature and humidity;

S3. Convert the wearing prediction result into two situations of wearing and not wearing according to the preset threshold, combine the wearing situations of each safety equipment, and divide the dangerous behavior into 8 levels based on the combination result;

According to the combined result of the edge detection confidence C _border and the personnel position detection confidence C _location , the grades of 4 objects and the construction environment are divided;

Determine the category of abnormal behavior according to the prediction results of personnel behavior and the proportion of each abnormal behavior in the dangerous behavior data set C _Db ;

同时实施预警。S4. Calculate the basic safety score SP, the environmental safety score _{S E} and the global safety score S _G of the personnel, and dynamically adjust the activation parameters of each safety score, so as to realize the dynamic adjustment of the warning level and obtain the warning score

At the same time implement early warning.

Further, step S1 is specifically:

Machine learning algorithm is used to clean the collected data, delete redundant data, and mark the historical data C _old to be marked to obtain the helmet label L _hat , seat belt label L _belt , mask label L _mask and construction vehicle label L _car , the dataset C is converted into a label dataset _CL ;

The data collected by the terminal equipment in the dangerous area and the data with dangerous behaviors in the data set C are marked respectively, and the dangerous area data set C _Da and the dangerous behavior data set C _Db are obtained.

Further, step S2 is specifically:

S201. Use the YOLO deep learning network to train and model the label data set _CL to obtain a safety wearing detection model, and simultaneously predict the wearing of the helmet, seat belt and mask in the same scene, and obtain the helmet wearing prediction score P _helmet , the seat belt wearing prediction score P _belt and the mask wearing prediction score P _mask ;

S202. Perform grayscale, filtering and edge detection processing on the data in the dangerous area data set C _Da , and then use Hough straight line transformation for straight line boundaries, use sliding windows for curve boundaries, and use irregular contours for irregular areas. The rotating rectangle replaces the irregular outline, thereby determining the danger area D _area and calculating the edge detection confidence C _border ; calculating the personnel position detection confidence C _location by the adjacent frame relationship and the obtained person position of the safety wear detection model;

将正常行为数据集

中第i个原始姿势Pi进行归一化，作为空间信息；通过对正常行为数据集

进行计算获取第i个骨架关节点的运动速度v_i，使用Rayleigh分布将离散分布的v_i建模为概率分布，即为人体的正常行为概率分布，将人体的正常行为概率分布作为时间信息；从实时采集的新数据C_new中生成历史运动图像MHI；S203. Deduct the data set C from the dangerous behavior data set C _Db to obtain the normal behavior data set

normal behavior dataset

The i-th original pose Pi is normalized as spatial information; by comparing the normal behavior data set

Perform calculation to obtain the motion speed vi of the _i - _th skeleton joint point, and use Rayleigh distribution to model the discrete distribution vi as a probability distribution, that is, the normal behavior probability distribution of the human body, and use the normal behavior probability distribution of the human body as time information; Generate historical moving image MHI from new data C _new collected in real time;

Using temporal information and spatial information, the regularity of human pose trajectory learned by the Transformer model based on space and time is constructed, and historical moving image MHI and non-negative sparse auto-encoder are introduced to make the Transformer model learn in a self-supervised incremental manner, The trained Transformer model is used as an abnormal behavior detection model to predict, and the abnormal behavior score F is obtained;

S204 , based on the mean value and standard deviation of the text data C _text of the ambient temperature and humidity, using the SPC statistical process discrimination standard, determine the data in the uncontrolled state in the batch, and calculate the data abnormality index E.

Further, the calculation process of the edge detection confidence C _border and the object position confidence C _location in S202 is as follows:

Among them, k ₁ is the filtering influence factor; kernel_size is the size of the Gaussian kernel used; k ₂ is the edge detection influence factor; t is the threshold of edge detection; h _SUB is the minimum contour height of the object obtained by the difference between adjacent frames; cur _h is the predicted height obtained by the safety wear detection model; β is the height scale factor.

Further, the personnel movement speed v _i in S203 is:

Among them, ( _{xi, j} , _{yi, j} ) and ( _{xi+1, j} , _{yi+1, j} ) represent the positions of the j-th joint point at the i-th frame and the i+1-th frame, respectively. 6. The construction environment-oriented multi-level security early warning method according to claim 3, wherein the data abnormality index E in S204 is calculated as follows:

N _normal represents the number of normal data, and N _abnormal represents the number of data in an uncontrolled state.

Further, step S3 is specifically:

S301. Combine the wear prediction results according to the importance of different wearable devices in construction, and divide them into 8 different levels of dangerous behaviors, as follows:

代表未佩戴安全帽；w_h代表佩戴安全带，

代表未佩戴安全带；w_m代表佩戴口罩，

代表未佩戴口罩，w_h、w_b、w_m∈{0，1}；Among them, w _h represents wearing a helmet,

Represents not wearing a helmet; w _h represents wearing a seat belt,

Represents not wearing a seat belt; w _m represents wearing a mask,

represents not wearing a mask, w _h , w _b , w _m ∈ {0, 1};

S302, determine the location coordinates of the personnel and the boundary coordinates of the D _area to obtain the initial out-of-bounds behavior; divide the boundary determination factor S _b according to the size of the edge detection confidence C _border , and divide according to the size of the personnel position detection confidence C _location The location determination factor S _l , based on the combination of the boundary determination factor S _b and the location determination factor S _l , divides the initial out-of-bounds behavior into four types of out-of-bounds confidence levels B _i , as follows:

The boundary determination factor S _b , the location determination factor S _l and the out-of-bounds behavior B _i are calculated as follows:

S303: Calculate the probability of the three types of abnormal behaviors of personnel in the dangerous behavior data set C _Db , such as slapsticking, falling and talking, and determine the abnormal behavior category A _j based on the abnormal behavior probability and the abnormal behavior score F, where A _j is:

Among them, c _fight represents the number of abnormal behaviors in the dangerous behavior dataset C _Db ; c _fall represents the number of abnormal falling behaviors in the dangerous behavior dataset C _Db ; c _sneak represents the abnormal behavior of talking in the dangerous behavior dataset C _Db The number in; F stands for abnormal behavior score;

S304. Set an alarm threshold T _alarm , compare the data abnormality index with the alarm threshold T _alarm , and give an early warning to the abnormal state. The definition of the abnormal state is as follows:

Among them, _Sk represents the kth state.

的计算方法为：Further, in step S4, the basic safety score SP of the personnel, the environmental safety score _{S E} , the global safety score S _G and the early warning score

The calculation method is:

S _P =D ₁ *P _helmet +D ₂ *P _belt +D ₃ *P _mask

S _E =C _border *C _location

S _G = F

Among them, D ₁ , D ₂ , D ₃ are risk factors; F is abnormal behavior score; a ₁ , a ₂ , a ₃ are activation parameters, a ₁ , a ₂ , a ₃ ∈ {0, 1}; w ₁ , w ₂ , w ₃ are detection factors; S _i is the activation parameter condition under the ith warning level.

Further, the risk factors D ₁ , D ₂ and D ₃ are:

Among them, _Ni represents the statistical number of _Li behavior levels in the labeled dataset _CL .

Compared with the prior art, the present invention has the following beneficial effects:

The invention proposes a multi-level security early warning method oriented to the construction environment, and adopts a hierarchical early-warning mechanism based on the difference between construction projects and general projects, which improves the systematicity, accuracy and reliability of the multi-level security early warning of construction big data; Labeling and hazard classification of the collected data, through in-depth mining of the correlation between multi-dimensional and multi-element data, under the premise of ensuring accuracy and timeliness, effectively saving the cost of early warning; in the detection process, combined with existing The advanced algorithm proposes a visual depth analysis scheme suitable for the project construction scene. The present invention effectively prevents the occurrence of dead ends in safety detection, and gives the system the ability to dynamically adjust the warning level.

Further, in step S1, processing operations such as labeling and hazard classification are performed through the collected real-time data on the construction site, so that the original data can provide richer semantic information in the multi-level early warning, which is convenient for subsequent operation and processing, and the data is completed. perception and collection.

Further, step S2 focuses on the semantics of the original data, and processes and analyzes different label data sets by combining advanced algorithms with traditional computer vision technology, so as to realize the extraction of key features of the data and the discrimination of basic information, which greatly reduces the cost of data processing. Early warning cost of construction big data.

Further, step S3 focuses on the three major aspects of the standardization of construction personnel's wearing, the relationship between objects and the environment, and the distribution of personnel behavior, while taking into account the real-time status of materials on the construction site, constructing early warning behaviors, and providing rule guarantees for multi-level early warning. At the same time, the early warning capability is further improved.

Further, in step S4, the system early warning score is calculated according to the calculation personnel safety, environmental safety and global safety scores, and the early warning level is dynamically adjusted, which can effectively adapt to various occasions and prevent the early warning from appearing in the three-level early warning of dead ends. Normal distribution, combined with human skeleton information, can effectively identify a variety of abnormal behaviors. By judging the rationality of human behavior in the whole, it can realize early warning of abnormal behavior and effectively prevent dead ends in early warning.

To sum up, the early warning method of the present invention fully realizes a multi-level safety early warning mechanism suitable for different construction scenarios, and involves a three-level early warning module systematically. The utilization rate of construction big data is improved, the adaptability of the model is improved through incremental learning, the cost of early warning is reduced through visual realization, and the multi-level early warning mechanism is used to effectively prevent the occurrence of dead ends in early warning.

Description of drawings

Fig. 1 is the flow chart of the multi-level three-dimensional project construction safety warning method of the present invention applied to industrial big data;

FIG. 2 is a schematic diagram of an end-edge-cloud architecture of the present invention;

Figure 3 is an implementation effect diagram.

Detailed ways

In order to make those skilled in the art better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only Embodiments are part of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

It should be noted that the terms "first", "second" and the like in the description and claims of the present invention and the above drawings are used to distinguish similar objects, and are not necessarily used to describe a specific sequence or sequence. It is to be understood that the data so used may be interchanged under appropriate circumstances such that the embodiments of the invention described herein can be practiced in sequences other than those illustrated or described herein. Furthermore, the terms "comprising" and "having" and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those expressly listed Rather, those steps or units may include other steps or units not expressly listed or inherent to these processes, methods, products or devices.

Below in conjunction with accompanying drawing, the present invention is described in further detail:

The invention provides a construction environment-oriented multi-level security early warning method, which performs correlation and hierarchical analysis on the original data, designs a multi-level early warning method based on industrial big data, and designs a multi-level early warning module for the original data collected by edge intelligent equipment. , firstly predict the normative behavior of personnel to achieve first-level early warning; for complex construction sites, by analyzing the relationship between the original data and the construction environment, to achieve the purpose of second-level early warning; for the overall state, three-dimensional analysis of personnel anomalies Sexual behavior, so as to achieve three-level early warning for construction big data.

Please refer to Figure 1. Figure 1 is a flowchart of a multi-level security early warning method oriented to a construction environment, a multi-level security early warning method oriented to a construction environment. The specific steps are as follows:

S1. Collect multi-dimensional construction big data C, the multi-dimensional construction big data C includes image video and text data C _text of ambient temperature and humidity, and the image video is composed of historical data C _old to be marked and new data collected in real time C _new composition;

The historical data C _old to be labeled is labeled in different dimensions to obtain a labeled data set _CL ; the labeled data set _CL includes a dangerous area data set C _Da and a dangerous behavior data set C _Db ;

Step S1 is specifically:

S101. Use a machine learning algorithm to clean the collected data, delete redundant data, and mark the historical data C _old to be marked to obtain a helmet label L _hat , a seat belt label L _belt , a mask label L _mask and a construction vehicle label L _car , the dataset C is converted into a label dataset _CL ;

The data collected by the terminal equipment in the dangerous area and the data with dangerous behaviors in the data set C are marked respectively, and the dangerous area data set C _Da and the dangerous behavior data set C _Db are obtained.

S2. Based on the label data set C _L , use the neural network to predict the wearing condition of the safety equipment, and obtain the wearing prediction result;

The dangerous area D _area is divided based on the dangerous area data set C _Da , and the edge detection confidence C _border is obtained by using the relevant parameters; the personnel position information is predicted by using the neural network, and the personnel position detection confidence C _location is obtained;

The historical moving image MHI is generated based on the _new data Cnew collected in real time, the spatiotemporal information of normal human movement is obtained based on the dangerous behavior data set C _Db , and the abnormal behavior detection model is constructed and updated by using the neural network based on the spatiotemporal information and the historical moving image MHI. Predict human behavior;

Calculate the mean and standard deviation of the data according to the text data C _text of ambient temperature and humidity;

Step S2 is specifically:

S201. Use the YOLO deep learning network to train and model the label data set _CL to obtain a safety wearing detection model, and simultaneously predict the wearing of the helmet, seat belt and mask in the same scene, and obtain the helmet wearing prediction score P _helmet , the seat belt wearing prediction score P _belt and the mask wearing prediction score P _mask .

S202. Perform grayscale, filtering and edge detection processing on the data in the dangerous area data set C _Da , and then use Hough straight line transformation for straight line boundaries, use sliding windows for curve boundaries, and use irregular contours for irregular areas. Rotate the rectangle to replace the irregular contour, so as to determine the dangerous area D _area , and use the parameter with strong correlation to calculate the edge detection confidence C _border , in addition, calculate the personnel position detection confidence through the relationship between adjacent frames and the person position obtained by the safety wear detection model C _location .

Further, the calculation process of the edge detection confidence C _border and the object position confidence C _location is as follows:

Among them, k ₁ is the filtering influence factor; kernel_size is the size of the Gaussian kernel used; k ₂ is the edge detection influence factor; t is the threshold of edge detection; h _SUB is the minimum contour height of the object obtained by the difference between adjacent frames; cur _h is the predicted height obtained by the safety wear detection model; β is the height scale factor.

将正常行为数据集

中第i个原始姿势Pi进行归一化，作为空间信息；通过对正常行为数据集

进行计算获取第i个骨架关节点的运动速度v_i，使用Rayleigh分布将离散分布的v_i建模为概率分布，即为人体的正常行为概率分布，将人体的正常行为概率分布作为时间信息；从实时采集的新数据C_new中生成历史运动图像MHI；S203. Deduct the data set C from the dangerous behavior data set C _Db to obtain the normal behavior data set

normal behavior dataset

The i-th original pose Pi is normalized as spatial information; by comparing the normal behavior data set

Perform calculation to obtain the motion speed vi of the _i - _th skeleton joint point, and use Rayleigh distribution to model the discrete distribution vi as a probability distribution, that is, the normal behavior probability distribution of the human body, and use the normal behavior probability distribution of the human body as time information; Generate historical moving image MHI from new data C _new collected in real time;

Using temporal information and spatial information, the regularity of human pose trajectory learned by the Transformer model based on space and time is constructed, and historical moving image MHI and non-negative sparse auto-encoder are introduced to make the Transformer model learn in a self-supervised incremental manner, The trained Transformer model is used as an abnormal behavior detection model for prediction, and the abnormal behavior score F is obtained.

Further, the calculation process of the personnel movement speed _vi is as follows:

Wherein, ( _xi,j , _yi,j ) and ( _xi+1,j , _yi+1,j ) represent the position of the jth joint point in the ith frame and the i+1th frame respectively;

S204 , based on the mean value and standard deviation of the text data C _text of the ambient temperature and humidity, using the SPC statistical process discrimination standard, determine the data in the uncontrolled state in the batch, and calculate the data abnormality index E.

Further, the data anomaly index E is calculated as follows:

Among them, N _normal represents the number of normal data, and N _abnormal represents the number of data in an uncontrolled state.

S3. Convert the wearing prediction result into two situations of wearing and not wearing according to the preset threshold, combine the wearing situations of each safety equipment, and divide the dangerous behavior into 8 levels based on the combination result;

According to the combined result of the edge detection confidence C _border and the personnel position detection confidence C _location , the grades of 4 objects and the construction environment are divided;

Determine the category of abnormal behavior according to the prediction results of personnel behavior and the proportion of each abnormal behavior in the dangerous behavior data set C _Db ;

Step S3 is specifically:

S301. Combine the wear prediction results according to the importance of different wearable devices in construction, and divide them into 8 different levels of dangerous behaviors, as follows:

代表未佩戴安全帽；w_h代表佩戴安全带，

代表未佩戴安全带；w_m代表佩戴口罩，

代表未佩戴口罩，w_h、w_b、w_m∈{0，1}；Among them, w _h represents wearing a helmet,

Represents not wearing a helmet; w _h represents wearing a seat belt,

Represents not wearing a seat belt; w _m represents wearing a mask,

represents not wearing a mask, w _h , w _b , w _m ∈ {0, 1};

S302, determine the location coordinates of the personnel and the boundary coordinates of the D _area to obtain the initial out-of-bounds behavior; divide the boundary determination factor S _b according to the size of the edge detection confidence C _border , and divide according to the size of the personnel position detection confidence C _location The location determination factor S _l , based on the combination of the boundary determination factor S _b and the location determination factor S _l , divides the initial out-of-bounds behavior into four types of out-of-bounds confidence levels B _i , as follows:

The boundary determination factor S _b , the location determination factor S _l and the out-of-bounds behavior B _i are calculated as follows:

S303: Calculate the probability of the three types of abnormal behaviors of personnel in the dangerous behavior data set C _Db , such as slapsticking, falling and talking, and determine the abnormal behavior category A _j based on the abnormal behavior probability and the abnormal behavior score F, where A _j is:

Among them, c _fight represents the number of abnormal behaviors in the dangerous behavior dataset C _Db ; c _fall represents the number of abnormal falling behaviors in the dangerous behavior dataset C _Db ; c _sneak represents the abnormal behavior of talking in the dangerous behavior dataset C _Db The number in; F stands for abnormal behavior score;

S304. Set an alarm threshold T _alarm , compare the data abnormality index with the alarm threshold T _alarm , and give an early warning to the abnormal state. The definition of the abnormal state is as follows:

Among them, _Sk represents the kth state.

同时实施预警。S4. Calculate the basic safety score SP, the environmental safety score _{S E} and the global safety score S _G of the personnel, and dynamically adjust the activation parameters of each safety score, so as to realize the dynamic adjustment of the warning level and obtain the warning score

At the same time implement early warning.

Step S4 is specifically:

的计算方法为：S401, personnel basic safety score _SP , environmental safety score _SE , global safety score _SG and early warning score

The calculation method is:

S _P =D ₁ *P _helmet +D ₂ *P _belt +D ₃ *P _mask

S _E =C _border *C _location

S _G = F

Among them, D ₁ , D ₂ , D ₃ are risk factors; F is abnormal behavior score; a ₁ , a ₂ , a ₃ are activation parameters, a ₁ , a ₂ , a ₃ ∈ {0, 1}; w ₁ , w ₂ , w ₃ are detection factors; S _i is the activation parameter condition under the ith warning level.

Further, the calculation process of D ₁ , D ₂ , and D ₃ is as follows:

Among them, Ni represents the statistical number of L_i behavior levels in the labeled dataset C_L.

To sum up, using the multi-level security early warning method oriented to the construction environment proposed by the present invention can continuously use newly acquired data to improve model capabilities, and at the same time, with the help of the three-level early warning method, multi-scene, three-dimensional and efficient dynamic early warning can be realized .

Referring to Figure 2, Figure 2 is an end-edge-cloud architecture model diagram, with data perception and collection-aggregation and integration-analysis and reasoning-early warning as the system, and the "end-edge- The cloud" model is the architecture, with regional autonomy and core early warning analysis and regulation as the design idea, mainly composed of three-level early warning to form a new three-dimensional security early warning mechanism. standardization, realize first-level early warning; and then analyze the relationship between safety-related native data and construction environment for complex construction site environment, such as cross-border detection of personnel and vehicles, joint detection of terminal temperature and humidity intelligent sensor data, etc., to achieve second-level Finally, starting from the overall situation, three-dimensionally analyze the abnormal behavior of personnel, establish a rule base for abnormal behavior in construction scenarios, and use the incremental learning method to define abnormal behavior according to the normal behavior probability distribution of skeletal motion statistics, so as to achieve industrial-scale Data-based project construction safety three-level early warning mechanism. At the same time, in order to overcome the high complexity, wide range, wide distribution and many types of construction scenarios, we research and design a three-dimensional security early warning “end-edge-cloud” architecture composed of intelligent terminal equipment that is autonomous in a small area and controlled by the cloud as a whole. The model always adheres to the security tenet of "people-oriented", and is specifically developed as the terminal intelligent device (end) completes the perception and collection of construction data, the edge server (side) completes data cleaning and data preprocessing and preliminary early warning of main label labeling , the cloud (cloud) realizes the establishment of multi-level early warning model rule base and dynamic early warning.

Example

The present invention is effective in construction site construction and railway construction scenarios. As shown in Figure 3, the first column is the original picture collected from the intelligent terminal equipment, the second column is the data representation obtained through the detection model, and the third column is the data representation obtained by the detection model. Columns are qualitative descriptions obtained after passing through the rule base.

The above content is only to illustrate the technical idea of the present invention, and cannot limit the protection scope of the present invention. Any changes made on the basis of the technical solution according to the technical idea proposed by the present invention all fall within the scope of the claims of the present invention. within the scope of protection.

Claims (9)

1. a multi-level security early warning method oriented to construction environment, is characterized in that, comprises the following steps; S1. Collect multi-dimensional construction big data C, the multi-dimensional construction big data C includes image video and text data C _text of ambient temperature and humidity, and the image video is composed of historical data C _old to be marked and new data collected in real time C _new composition; The historical data C _old to be labeled is labeled in different dimensions to obtain a labeled data set _CL ; the labeled data set _CL includes a dangerous area data set C _Da and a dangerous behavior data set C _Db ; S2. Based on the label data set C _L , use the neural network to predict the wearing condition of the safety equipment, and obtain the wearing prediction result; The dangerous area D _area is divided based on the dangerous area data set C _Da , and the edge detection confidence C _border is obtained by calculation; the personnel position information is predicted by using the neural network, and the personnel position detection confidence C _location is obtained; The historical moving image MHI is generated based on the _new data Cnew collected in real time, the spatiotemporal information of normal human movement is obtained based on the dangerous behavior data set C _Db , and the abnormal behavior detection model is constructed and updated by using the neural network based on the spatiotemporal information and the historical moving image MHI. Predict human behavior; Calculate the mean and standard deviation of the data according to the text data C _text of ambient temperature and humidity; S3. Convert the wearing prediction result into two situations of wearing and not wearing according to the preset threshold, combine the wearing situations of each safety equipment, and divide the dangerous behavior into 8 levels based on the combination result; According to the combined result of the edge detection confidence C _border and the personnel position detection confidence C _location , the grades of 4 objects and the construction environment are divided; Determine the category of abnormal behavior according to the prediction results of personnel behavior and the proportion of each abnormal behavior in the dangerous behavior data set C _Db ; S4. Calculate the basic safety score SP, the environmental safety score _{S E} and the global safety score S _G of the personnel, and dynamically adjust the activation parameters of each safety score, so as to realize the dynamic adjustment of the warning level and obtain the warning score

Also implement early warning. 2. The multi-level safety early warning method oriented to construction environment according to claim 1, is characterized in that, step S1 is specifically: Machine learning algorithm is used to clean the collected data, delete redundant data, and mark the historical data C _old to be marked to obtain the helmet label L _hat , seat belt label L _belt , mask label L _mask and construction vehicle label L _car , the dataset C is converted into a label dataset _CL ; The data collected by the terminal equipment in the dangerous area and the data with dangerous behaviors in the data set C are marked respectively, and the dangerous area data set C _Da and the dangerous behavior data set C _Db are obtained. 3. the construction environment-oriented multi-level safety early warning method according to claim 1, is characterized in that, step S2 is specifically: S201. Use the YOLO deep learning network to train and model the label data set _CL to obtain a safety wearing detection model, and simultaneously predict the wearing of the helmet, seat belt and mask in the same scene, and obtain the helmet wearing prediction score P _helmet , the seat belt wearing prediction score P _belt and the mask wearing prediction score P _mask ; S202. Perform grayscale, filtering and edge detection processing on the data in the dangerous area data set C _Da , and then use Hough straight line transformation for straight line boundaries, use sliding windows for curve boundaries, and use irregular contours for irregular areas. The rotating rectangle replaces the irregular outline, thereby determining the danger area D _area and calculating the edge detection confidence C _border ; calculating the personnel position detection confidence C _location by the adjacent frame relationship and the obtained person position of the safety wear detection model; S203. Deduct the data set C from the dangerous behavior data set C _Db to obtain the normal behavior data set

normal behavior dataset

The i-th original pose Pi is normalized as spatial information; by comparing the normal behavior data set

Perform calculation to obtain the motion speed vi of the _i - _th skeleton joint point, and use Rayleigh distribution to model the discrete distribution vi as a probability distribution, that is, the normal behavior probability distribution of the human body, and use the normal behavior probability distribution of the human body as time information; Generate historical moving image MHI from new data C _new collected in real time; Using temporal information and spatial information, the regularity of human pose trajectory learned by the Transformer model based on space and time is constructed, and historical moving image MHI and non-negative sparse auto-encoder are introduced to make the Transformer model learn in a self-supervised incremental manner, The trained Transformer model is used as an abnormal behavior detection model to predict, and the abnormal behavior score F is obtained; S204 , based on the mean value and standard deviation of the text data C _text of the ambient temperature and humidity, using the SPC statistical process discrimination standard, determine the data in the uncontrolled state in the batch, and calculate the data abnormality index E. 4. the multi-level security early warning method oriented to construction environment according to claim 3, is characterized in that, in S202, the calculation process of edge detection confidence C _border and object position confidence C _location is as follows:

Among them, k ₁ is the filtering influence factor; kernel_size is the size of the Gaussian kernel used; k ₂ is the edge detection influence factor; t is the threshold of edge detection; h _SUB is the minimum contour height of the object obtained by the difference between adjacent frames; cur _h is the predicted height obtained by the safety wear detection model; β is the height scale factor. 5. the construction environment-oriented multi-level safety early warning method according to claim 3, is characterized in that, the personnel movement speed v _i in S203 is:

Among them, ( _{xi, j} , _{yi, j} ) and ( _{xi+1, j} , _{yi+1, j} ) represent the positions of the j-th joint point at the i-th frame and the i+1-th frame, respectively. 6. the construction environment-oriented multi-level safety early warning method according to claim 3, is characterized in that, the data abnormality index E in S204 is calculated as follows:

N _normal represents the number of normal data, and N _abnormal represents the number of data in an uncontrolled state. 7. The multi-level safety early warning method oriented to construction environment according to claim 3, is characterized in that, step S3 is specifically: S301. Combine the wear prediction results according to the importance of different wearable devices in construction, and divide them into 8 different levels of dangerous behaviors, as follows:

Among them, w _h represents wearing a helmet,

Represents not wearing a helmet; w _h represents wearing a seat belt,

Represents not wearing a seat belt; w _m represents wearing a mask,

represents not wearing a mask, w _h , w _b , w _m ∈ {0, 1}; S302, determine the location coordinates of the personnel and the boundary coordinates of the D _area to obtain the initial out-of-bounds behavior; divide the boundary determination factor S _b according to the size of the edge detection confidence C _border , and divide according to the size of the personnel position detection confidence C _location The location determination factor S _l , based on the combination of the boundary determination factor S _b and the location determination factor S _l , divides the initial out-of-bounds behavior into four types of out-of-bounds confidence levels B _i , as follows: The boundary determination factor S _b , the location determination factor S _l and the out-of-bounds behavior B _i are calculated as follows:

S303: Calculate the probability of the three types of abnormal behaviors of the personnel in the dangerous behavior data set C _Db , such as slapsticking, falling, and talking, and determine the abnormal behavior category A _j based on the abnormal behavior probability and the abnormal behavior score F, where A _j is:

Among them, c _fight represents the number of abnormal behaviors in the dangerous behavior dataset C _Db ; c _fall represents the number of abnormal falling behaviors in the dangerous behavior dataset C _Db ; c _sneak represents the abnormal behavior of chatting in the dangerous behavior dataset C _Db The number in; F represents the abnormal behavior score; S304. Set an alarm threshold T _alarm , compare the data abnormality index with the alarm threshold T _alarm , and give an early warning to the abnormal state. The abnormal state is defined as follows:

Among them, _Sk represents the kth state. 8. The construction environment-oriented multi-level safety early warning method according to claim 1, characterized in that, in step S4, personnel basic safety score _SP , environmental safety score _SE , global safety score _SG and early warning score

The calculation method is: S _P =D ₁ *P _helmet +D ₂ *P _belt +D ₃ *P _mask S _E =C _border *C _location S _G = F

Among them, D ₁ , D ₂ , D ₃ are risk factors; F is abnormal behavior score; a ₁ , a ₂ , a ₃ are activation parameters, a ₁ , a ₂ , a ₃ ∈ {0, 1}; w ₁ , w ₂ , w ₃ are detection factors; S _i is the activation parameter condition under the ith warning level. 9. The construction environment-oriented multi-level safety early warning method according to claim 8, wherein the risk factors D ₁ , D ₂ and D ₃ are:

Among them, _Ni represents the statistical number of _Li behavior levels in the labeled dataset _CL .

Images