LU501401B1 - Action risk identification method for power system field operation based on graph convolution - Google Patents

Abstract

The present disclosure relates to a dynamic action risk identification method for power system field operators, the method extracts skeleton information of the field operators in the power system by using a human posture estimation method, thereby transforming the video information into an undirected graph that contains the skeleton information. Then a spatiotemporal graph convolution network is used to realize the action identification of the field operators. The present disclosure can realize the risk identification of the dynamic behaviors of the field operators. The violations and risks of the field operators can be identified in real time, thereby providing technical means for the early warning and management of safety risks at the power production site, and the probability of power system accidents can be reduced. Meanwhile, the intrinsic safety level of power production is improved.

Description

BL-5405

ACTION RISK IDENTIFICATION METHOD FOR POWER SYSTEM 4501401

FIELD OPERATION BASED ON GRAPH CONVOLUTION

TECHNICAL FIELD

[01] The present disclosure relates to a field of power system operation safety management and control, and more particularly, to a dynamic action risk identification method for power system field operators based on graph convolution.

BACKGROUND ART

[02] Safe production is the basic guarantee for stable operation of power system.

Once a security accident occurs in the power system, it will cause huge economic losses and adverse social impact. The real-time identification and control of on-site operation risk of the power system is of great significance to ensure the personal safety of operators and the safe and stable operation of power grid.

[03] At present, the safety risk management methods for on-site operation mainly include manual safety supervision method and video monitoring method. The manual safety supervision method is mainly to supervise the behavior and operation of operators by specially arranging supervisors. However, the supervisors cannot guarantee all-round supervision of the operators. Moreover, the supervisors are as susceptible to external factors as operators, their attention may be distracted, which may lead to safety accidents. In particular, the power system operation has the characteristics of large amount of on-site operations, large numbers of equipment and complex operation.

Therefore, the manual supervision method is unable to realize the real-time supervision and risk early warning of all operation processes. The video monitoring system provides effective assistance for safety supervision, but the actual monitoring task still needs more manual work. The monitoring system usually only records video images for later evidence collection. In addition, some scholars have proposed the intelligent analysis methods based on surveillance video, including personnel information verification of on-site operators, safety helmet and safety belt detection, transformer substation fire detection and billboard detection. The existing methods mainly focus on static safety risk identification. However, the actual field operation in power system is a continuous and dynamic process, in which there are many dynamic security risks and violations.

The existing safety risk identification methods for safety supervision cannot identify those dynamic risks and violations of operators.

[04] Therefore, the present disclosure introduces a risk identification method for dynamic behavior of field operators of power system, to realize the dynamic action risk identification, real-time safety management and control of the field operators.

SUMMARY

[05] The purpose of the present disclosure is to provide a risk identification method for dynamic behavior of field operators of power system, which can realize dynamic risk identification of field operators of power system. The main steps of the method are as follows:

[06] Step 1, behavior monitoring and image transmission of the field operators of a 1

BL-5405 power system. Monitoring behavior of the operators through monitoring cameras LU501401 arranged on operation site of the power system, and uploading monitoring videos to a central cloud platform.

[07] Step 2, operator behavior posture estimation. Estimating behavior postures of field operators in the monitoring videos by using openpose, a human posture estimation model , then obtaining skeleton information of the field operators' behavior.

[08] Step 3, connecting the obtained human skeleton information and constructing an undirected graph G=(v,A,X) containing action information of the field operators,

N wherein Jom is a set of N vertices, A is a weighted adjacency matrix, X is a signal matrix on the vertices, and a vertex set V corresponds to all N joints of the skeleton, the adjacency matrix A represents a link relationship between two vertices, wherein if vertex vi and vertex vj are directly connected, Ajj = 1, otherwise Ai = 0.

[09] Step 4, transforming the constructed undirected graph into a spectral domain, and filtering the undirected graph containing human skeleton information in the spectral domain. For an undirected graph G=(v,A,X), a Laplace matrix L of the undirected graph is defined as follows:

[10] L=D-A;

[11] wherein, D is a degree matrix of vertices, and diagonal elements are degrees of each vertex in turn.

[12] The Laplace matrix L of the undirected graph is a symmetric positive definite matrix. The Laplace matrix L can be decomposed as follows:

[13] L=UAU",

[14] Where, A=diag([A ‚A ‚...,An ]) is a diagonal matrix of eigenvalues, and U— [ui ,u2,...,un ] is an orthogonal matrix corresponding to eigenvalue vectors.

[15] Presetting a filtering function g (-) of the undirected graph G the frequency domain filtering response of input signal X can be defined as Z(A)=X(4)g(A4), and the inverse Fourier transform of graph is defined as follows: 16] 10=D FAA) ;

[17] therefore, the matrix description of undirected graph filtering can be defined as follows:

[18] Z=Uldiag[g(4),&(%),...,&(A,)DU'LX .

[19] Step 5, identifying the undirected graph containing human skeleton information by using a graph convolutional network (GCN).

[20] Step 6, adding a human key bone node attention module between the graph convolution networks, focusing on key bone nodes related to actions, to improve the feature expression and action recognition accuracy of the network. For the violation behaviors of the operators in the power system, the influence of each skeleton joint on action recognition is different. Because an action of the operators usually has a strong correlation with only a few bone joints, weak correlation with other joints. Therefore, the patent proposes to use a human key bone node attention module to enhance the 2

BL-5405 information of key bone joints, thus improve the accuracy of operator's illegal action LU501401 identification. Assuming the feature graph obtained by graph convolution is H, the output feature graph H 'obtained by the key bone node attention module is as follows.

[21] H=WWHH;

[22] where, Wj; >0 and SW, =1. The larger the Wj, i-th joint node is more important for violation identification.

[23] Step 7, action identification and classification. For feature graphs containing human posture information extracted by the graph convolution network, all feature graphs are converted into a one-dimensional matrix by using use a full connection layer, and then a softmax function is used to identify and classify illegal behaviors and risk actions of different operators.

[24] The features of the present disclosure are:

[25] The present disclosure converts the video information into the undirected graph containing the skeleton information by extracting the skeleton information of the field operators of the power system. Then uses the graph convolution network to realize the action identification of the field operator. The present disclosure can realize the risk identification of the dynamic behaviors of the field operators. It also can identify the violations and risks of the field operators in real time, thereby providing technical means for the early warning and management of safety risks at the power production site. Simultaneously, reducing the probability of power system accidents and improving the intrinsic safety level of the power production.

BRIEF DESCRIPTION OF THE DRAWINGS

[26] Fig. 1 is a flowchart of an action risk identification method for power system field operation based on graph convolution of the present disclosure.

DETAILED DESCRIPTION OF THE EMBODIMENTS

[27] The present disclosure will be further described in detail below in combination with the accompanying drawings and specific embodiments, but the present disclosure is not limited to these embodiments.

[28] Specific embodiment:

[29] A flowchart of an action risk identification method for power system field operation based on graph convolution is shown in Fig. 1, and the specific steps are as follows:

[30] Step 1, behavior monitoring and image transmission of field operators of a power system. Monitoring behavior of the operators through monitoring cameras arranged on operation site of the power system, and uploading monitoring videos to a central cloud platform.

[31] Step 2, operator behavior posture estimation. Estimating behavior postures of the field operators in the monitoring videos by using a human posture estimation model openpose, and obtaining skeleton information of the field operators’ behavior.

[32] Step 3, connecting the obtained human skeleton information to construct an undirected graph G=(v,A,X) containing action information of the field operators. 3

BL-5405

N LU501401 wherein U Jom is a set of N vertices, A is a weighted adjacency matrix, X is a signal matrix on the vertices, and a vertex set V corresponds to all N joints of the skeleton, the adjacency matrix A represents a link relationship between two vertices, wherein if vertex vi and vertex vj are directly connected, Aj = 1, otherwise Ajj = 0.

[33] Step 4, transforming the constructed undirected graph into a spectral domain, filtering the undirected graph which contained the human skeleton information in the spectral domain. For an undirected graph G=(v,A,X), a Laplace matrix L of the undirected graph is defined as follows:

[34] L=D-A;

[35] where D is a degree matrix of vertices, and diagonal elements are degrees of each vertex in turn.

[36] The Laplace matrix L of the undirected graph is a symmetric positive definite matrix. The Laplace matrix L can be decomposed as follows:

[37] L=UAU",

[38] WhereA—diag([M ‚A ‚...,An ]) is a diagonal matrix of eigenvalues, and U— [ui ,u2,...,un ] is an orthogonal matrix corresponding to eigenvalue vectors.

[39] Presetting a filtering function g (-) of the undirected graph G the frequency domain filtering response of input signal X can be defined as Z(A)=X(4)g(A4), and the inverse Fourier transform of graph is defined as follows:

Ol 20=D 509800 ;

[41] therefore, the matrix description of undirected graph filtering can be defined as follows:

[42] Z =Uldiag[g(4),&(%,),...,&(4,)DU'LX .

[43] Step 5, identifying the undirected graph containing human skeleton information by using a graph convolutional network (GCN).

[44] Step 5.1, for the obtained undirected graph G(V,E) of human body architecture, calculating the degree matrix D, adjacency matrix A and Laplace matrix L, then performing the graph convolution on the input undirected graph G to get an output result as shown below: 3 i as] Cod AD SEW.

[46] Where, | represents layer input, W is weight matrix and is quantity to be solved.

[47] Step 5.2, inputting the undirected graph G into four graph convolution neural network modules, respectively, Block] 3x3, Block2 64x3, Block3 128x3 and

Block4 256x3 for feature extraction.

[48] Step 6, adding a human key bone node attention module between the graph convolution networks. Focusing on key bone nodes related to actions, to improve the feature expression and action recognition accuracy of the network.

[49] For the violation behaviors of operators in the power system, the influence of each skeleton joint on action recognition is different. Due to an action of the operators 4

BL-5405 . . ne . LU501401 usually has a strong correlation with only a few bone joints, and it has a weak correlation with others. Therefore, the patent proposes to use a human key bone node attention module to enhance the information of key bone joints, to improve the accuracy of operator's illegal action identification. If the feature graph obtained by graph convolution is H, then the output feature graph H 'obtained by the key bone node attention module is as follows.

[50] H'=W(H)H;

[51] wherein, Wj >0 and SW, =1. The larger the W;;, i-th joint node is more important for violation identification.

[52] Step 7, action identification and classification. For feature graphs containing human posture information extracted by the graph convolution network, all feature graphs are converted into a one-dimensional matrix by using use a full connection layer with 256 neurons. Then a softmax function is used to identify and classify illegal behaviors and risk actions of different operators.

Claims (1)

BL-5405 WHAT IS CLAIMED IS: LUS01401

1. An action risk identification method for power system field operation based on graph convolution consists of the following steps: step 1, behavior monitoring and image transmission of field operators of a power system. Monitoring behaviors of the operators through monitoring cameras which arranged on the operation site of the power system. Then uploading monitoring videos to a central cloud platform. step 2, operator behavior posture estimation. Estimating behavior postures of field operators in the monitoring videos by using a human posture estimation model openpose, and obtaining the skeleton information of the field operators' behavior. step 3, constructing an undirected graph that contains the human skeleton information. Connecting the obtained human skeleton information and constructing an undirected graph G=(v,A,X) which contained the action information of the field N operators. where, Jum is a set of N vertices, A is a weighted adjacency matrix, X is a signal matrix on the vertices, and a vertex set V corresponds to all N joints of the skeleton. The adjacency matrix A represents a link relationship between two vertices, where, if vertex vi and vertex v; are directly connected, Ajj = 1, otherwise Ai = 0; step 4, transforming the constructed undirected graph into a spectral domain, and filtering the undirected graph that contains the human skeleton information in the spectral domain. step 5, identifying the undirected graph that contains the human skeleton information by using a spatiotemporal graph convolution network. step 6, adding a human key bone node attention module between the graph convolution networks. A feature graph obtained by graph convolution is H, and an output feature graph H 'obtained by the key bone node attention module is as follows: H'=W(H)H; wherein, Wj >0 and SW, =1, and the larger the W;;, i-th joint node is more important for violation identification. step 7, action identification and classification. For the feature graphs containing human posture information extracted by the graph convolution network, all feature graphs are converted into a one-dimensional matrix by using a full connection layer. Then a softmax function is used to identify and classify the illegal behaviors and risk actions of different operators. 1