A Power Construction Safety Hazard Identification Technology Based on Surveillance Video on Site

There are more and more newly established substations and transmission lines, contributing to subsequent and reliable power supply and socioeconomic development. Violations are prone to occur on construction sites arousing safety hazards. This paper describes a video-based approach to identifying safety risks on site, building typical security hazards and violations feature libraries, and developing a BP neural network algorithm to identify safety risks. That approach can effectively improve the efficiency of on-site safety inspections on power infrastructure sites and contribute to the compliance of power infrastructure operations.


Introduction
To guarantee work safety and avoid safety incidents in power construction projects, it is necessary to strengthen the safety audit on site.Potential and present violations should be stopped once found to ensure work safety.By installing video surveillance terminals such as monitoring ball prisms with a video-image conversion algorithm and security hazards and violations feature libraries, remote audit, and intelligent checks for potential safety hazards on site become possible, which greatly reduces the workload of on-site inspectors [1].This paper describes a construction safety hazard identification approach.By adopting this approach, the efficiency of on-site safety inspections on power infrastructure sites can be improved greatly.

The identification process for potential safety hazards on power infrastructure sites
Traditional construction inspections usually require workers inspecting on-site costing great labor with low efficiency.When the number of construction projects that need inspection is increasing, traditional inspection manners cannot meet their needs.It is necessary to utilize modern information technology for remote safety work audits and hazard identification for potential and present safety risks [2][3].The surveillance video terminal carrying the approach proposed in this paper proposed consists of an image capture module, hazard identification module, video record module, and so on.The key three steps in the process of hazard identification are: (1) to acquire on-site images, (2) to identify, analyze those images, and output outcomes, (3) to generate alert information for potential and present violation prevention.The process is shown in Figure 1.

Application of video surveillance system in the identification of potential safety hazards on power infrastructure site
Since there are many construction areas requiring inspections, surveillance video terminals should be installed on requisite every area.Image capture modules should use thermal cameras being waterproof and dust-proof that can perform well under different climate changes [4][5].The camera should have larger than 1080 P pixels and greater than 20 x zoom.After capturing images, the process enters the process of identification and analysis.

Typical safety hazards and violating behavior characteristics of power infrastructure construction sites
To realize intelligent judgment on construction safety hazards and violating behaviors, a corresponding feature library should be built first.By comparing the image captured with the information in the feature library, the algorithm judges whether safety hazards or violations occur or not.The typical feature library covers various cases such as fire-forbidden signs in necessary areas, alert signs for excavation work in necessary areas, workers should wear safety belts at heights, workers should wear safety helmets wherever in the construction area, and so on [6].

The application process of image object detection on a construction site
Once a typical feature library, features can be identified by adopting neural network technology and compared with features in the library.Neural network technology is a machine learning technology that simulates the human brain and can extract key features from captured images.When it is applied in the surveillance video terminals, it will optimize the brightness or contrast, prune, and compress images captured per five seconds accordingly [7].Then, it will extract features shown in those images such as alert signs, workers' behaviors and wearables, intrusion, and so on.The process of identification is shown in Figure 2. The neural network model training is the key, to determining the accuracy of feature extraction.A type of BP neural network algorithm is introduced in this paper showing good results in the pilot application.

On-site image feature extraction and training process of power infrastructure
The BP neural network has two working signals, namely the working signal and the error signal.
(1) The working signal is a function of the propagation relationship of the signal.It allows the relationship between the input value and the weight to be established, and then the signal propagation becomes the output signal output.
(2) Error signal, which is a signal that propagates in the back.In the process of implementing the operation of the network, there will be errors between the output signal and the actual value.The function of the error signal is to transmit this information back to the input and re-enter the network.
The network signals propagate from the input layer to the output layer.The network signals are analyzed by values and weights before entering hidden layers.The hidden layer has many neurons like brain synapses that function and process incoming signals.The output layer receives processed signals and outputs predicted results.If the results are far from the actual, a reverse transmission process is carried out [8].In the reverse transmission process, the parameters of signals will be changed and signals will re-enter the network.The BP network learns with a large amount of training.Identification errors eliminate after the training.The feature extraction process of the BP network is shown in Figure 3.
The input layer is . The specific steps of the execution process of the BP neural network are as follows.
(1) We initialize the network; (2) We input vector ( , ) k k X Y into the network; (3) The learning vector is input to the input layer, and if the input layer cannot handle it, it will be sent to the hidden layer; (4) The node input and output values of the hidden layer can be determined by the following equation: (5) The node input and output values of the output layer can be determined by the following equation.
) (7) The following Equation ( 6) is the correction error of the hidden layer: (8) We adjust the hidden layer weights and thresholds by the following types: where learning efficiency is (0,1) a ∈ ; (9) Due to the existence of prediction error, the weight threshold is corrected by the following equation; where learning rate is (0, . (10) We repeat Step (3) with the re-selected input vector until all data have been trained; (11) We determine whether the error reaches the expected value; (12) If the error reaches the standard, the learning process of the network ends.If the standard is not met, the parameters of the network are adjusted; (13) If the error is still not satisfied, we return to Step (2).After the above steps, the key features in the image of the power infrastructure construction site can be effectively extracted and further used to identify whether there are potential safety hazards at the power construction site.

The advantages and disadvantages of applying BP neural network to infrastructure site safety hazard identification
Now, the neural network algorithm is one of the main algorithms of artificial intelligence, which is widely used in image recognition.BP neural network algorithm has three main advantages.
(1) If the structure of the setting network is reasonable, the nonlinear mapping ability of the BP neural network is very strong; (2) BP neural networks have strong adaptive abilities; (3) BP neural networks are highly fault-tolerant; Although it has many advantages, there are still many disadvantages, such as: (1) BP neural networks take a long convergence time.BP neural networks reduce prediction errors by updating weights and thresholds, and the error surface distribution of two-dimensional weighted spaces is particularly complex, resulting in longer convergence times.
(2) BP neural networks are prone to fall into local minimums, and their error surfaces are uneven, like hills one after another.It has many minimum points, which will cause the network to fall into local minimum points.
(3) The structure of BP neural networks is very complex, and how determining an optimal structure is a difficult point.How to determine the number of hidden layer nodes is a key point, if the number of hidden layer nodes is too much, the performance of the network will be reduced, if too few, the network will not be able to meet the accuracy requirements [9].In real-world experiments, scholars will screen the hidden layer based on trial methods or experience.
(4) When training the samples of the BP neural network, the loss of old samples will occur; (5) The learning step of the BP neural network is difficult to determine, the speed of convergence depends on the step size of learning.If the learning step is too large, it will lead to oscillation, impacting the convergence of the neural network, sometimes even not converging.If the learning step is too small, the speed will be slower, and the convergence time will be prolonged.
With the comparison of its advantages and disadvantages, it is found that there are still some limits and more study needs to be done for better performance on construction safety inspection and audit.

Conclusion
The improvement of construction safety hazard prevention and safety audit conduction is of great significance to power construction and society.With the development of information and technology, information technology will be widely adopted in the field of remote inspection and safety audits onsite.The construction safety hazard identification technology proposed in this paper shows good results in the piloted application and is worth developing further for wider and better usage in various scenarios.

Figure 1 Figure 2 Figure 3
Figure 1 Identification process of safety hazards on power infrastructure sites