An automatic identification system for personnel behavior in power systems based on 5G MEC offload

Abstract

The present invention discloses a 5G MEC-divided power system personnel behavior automatic identification system, in which a wireless CPE receives a video signal transmitted on site, and forwards the video signal to a 5G base station, which is transmitted to a MEC edge computing server through a 5G network, avoiding the public network, and realizing the first diversion; then the MEC edge computing server performs a second diversion on the first diversion result to obtain a monitoring video stream and a personnel behavior stream, and the filtered monitoring video stream is sent to a central control room host, and the personnel behavior stream is sent to an algorithm server; the algorithm server performs moving target detection on the personnel behavior stream. By performing a secondary diversion on the data, different transmission paths and calculations are selected for different data, and data and bandwidth matching is realized, which can not only ensure the timeliness of transmission of different types of data streams, but also save equipment deployment costs.

Description

Technical Field

[0001] The present invention relates to the technical field of power system security monitoring, and specifically to a 5G MEC-diverted power system personnel behavior automatic identification system and method. Background Art

[0002] In recent years, emerging technologies such as artificial intelligence have developed rapidly, prompting the widespread application of intelligent algorithms such as behavior detection and target tracking. For example, drone inspections of cables can analyze and identify whether there are foreign objects in the cables, greatly reducing the possibility of power safety accidents.

[0003] With the development of 5G technology and MEC edge computing servers, traffic diversion technology has been applied in some scenarios. At present, the traffic diversion technology based on 5G MEC edge computing servers diverts the public network and private network (5G sinking) to achieve data security issues. However, no data processing is performed, and non-target data occupies a large amount of bandwidth, resulting in large calculation volume and high equipment deployment costs.

[0004] For example, a mobile edge computing system and device for video surveillance services disclosed in application number CN202010849583.0, in which the mobile edge computing system includes: video surveillance equipment, wireless terminals and mobile edge computing servers, the video surveillance equipment is connected to the mobile communication network through the wireless terminal, and the mobile edge computing server is deployed at the back end of the base station in the mobile communication network to provide mobile edge computing services; the wireless terminal is connected to the mobile communication network deployed with the mobile edge computing server, and after the wireless terminal is connected to the communication network, the mobile edge computing server diverts the local monitoring video content according to the diversion rules and stores it in the local content server, and the mobile edge computing server also processes the local video through the pre-configured video recognition and processing strategy when diverting. The invention can quickly realize calculations by deploying local edge computing, and then make decisions and controls, with strong real-time performance, but the video signals collected by the front end are currently transmitted through the public network, which has transmission delays and cyberspace security issues.

[0005] According to the above discussion, the invention technology only completes the diversion of public and private networks, and does not perform secondary diversion of data. There are still problems such as large amount of calculation, large bandwidth occupancy, and high equipment deployment cost. Summary of the invention

[0006] The technical problem to be solved by the present invention is that the existing diversion technology only completes the diversion between the public network and the private network, and its application in the identification of personnel safety behavior in the power system still has the problems of large calculation amount and large bandwidth occupancy.

[0007] The present invention solves the above technical problems through the following technical means:

[0008] A 5G MEC-diverted power system personnel behavior automatic identification system, on-site video acquisition equipment, wireless CPE, 5G base station, MEC edge computing server, algorithm server, operator station host; the on-site video acquisition equipment is communicatively connected to the wireless CPE, the wireless CPE is communicatively connected to the 5G base station, the 5G base station is communicatively connected to the MEC edge computing server; the MEC edge computing server is communicatively connected to the algorithm server and the operator station host respectively; the algorithm server is communicatively connected to the operator station host;

[0009] The specific process of video signal diversion is as follows:

[0010] The wireless CPE receives the video signal transmitted on site, and forwards the video signal to the 5G base station, which is then transmitted to the MEC edge computing server through the 5G network, avoiding the public network and achieving the first diversion. The MEC edge computing server then performs a second diversion on the first diversion result to obtain the monitoring video stream and the personnel behavior stream, and transmits the diversion result to the firewall for filtering. The filtered monitoring video stream is sent to the host in the central control room, and the personnel behavior stream is sent to the algorithm server. The algorithm server performs motion target detection on the personnel behavior stream, identifies the personnel motion area in the video frame, and extracts features from the motion area. The extracted features are used as samples for training and modeling in the neural network algorithm, and finally completes the classification and recognition of personnel motion behavior.

[0011] The system provided by the present invention divides video information into monitoring video stream and personnel behavior stream through the 5G MEC edge computing server. Target recognition is only for the personnel behavior stream, and the amount of calculation is greatly reduced. Different transmission paths and adapted optical fibers are deployed according to different information flows, which can minimize the optical fiber cost and data processing volume, and achieve data and bandwidth matching, which can not only ensure the timeliness of transmission of different types of data streams, but also save equipment deployment costs.

[0012] Furthermore, the MEC edge computing server is configured with a diversion structure for implementing the second diversion, specifically:

[0013] The MEC edge computing server is configured with access and mobility management unit AMF, session management unit SMF, user plane function unit UPF, trusted domain PCF, untrusted domain NEF, edge computing service function unit MEC AF, service function request unit AF Request, user plane function unit UPF (UL CL), terminal UE, access network RAN, anchor point 1 UPF anchor1, anchor point 2 UPF anchor2;

[0014] The terminal UE is communicated with the MEC AF, the MEC AF is communicated with the NEF, the NEF is communicated with the PCF, the NEF and PCF are communicated with the AMF and SMF respectively, the AMF is communicated with the UE and the RAN respectively, the UE is communicated with the RAN, the SMF and the RAN are communicated with the UPF (UL-CL) respectively, and the UPF (UL-CL) is provided with two anchor points UPF anchor1 and UPF anchor2. The anchor points UPFanchor1 and UPF anchor2 connect the monitoring video stream and the personnel behavior stream to two different paths of the same data network.

[0015] Furthermore, the front-end video acquisition equipment, wireless CPE, and 5G base station are deployed on site; there are multiple front-end video acquisition devices, and a wireless CPE is deployed between two adjacent front-end video acquisition devices; and the 5G base station is deployed within 3-5m from the wireless CPE.

[0016] Furthermore, the MEC edge computing server and algorithm server are arranged in the communication room.

[0017] Furthermore, an isolation network gateway is also arranged in the communication room; the 5G base station is connected to the isolation network gateway through type A optical fiber; the isolation network gateway is connected to the switch through type A optical fiber communication, and the switch is connected to the MEC edge computing server through type A optical fiber; a bastion host is deployed in the bypass of the switch; after the MEC edge computing server diverts the signal, the on-site real-time monitoring video stream is transmitted to the operator station host based on type A optical fiber through the Winut firewall, and the operator station host is located in the control room; the MEC edge computing server transmits the personnel behavior video stream based on type B optical fiber through the Winut firewall to the algorithm server for calculation and identification, and the bandwidth of type B optical fiber and type A optical fiber will be selected according to the uplink limit of the amount of transmitted data; the algorithm server will be connected to the operator station host through type B optical fiber.

[0018] Furthermore, the algorithm server is configured with a power system personnel safety behavior recognition program, and the specific recognition process is:

[0019] Step 1. Obtain video information of power workers' operation behavior;

[0020] Step 2. Use the VIBE algorithm to detect the motion area in the video information and obtain the motion area;

[0021] Step 3. Use Itti algorithm to extract motion area features, including brightness features, color features, direction features, texture features, and shape features; among them, texture features The calculation method is:

[0022] 1) For sports areas ,make and Represent the horizontal and vertical gradients of the motion area respectively, then the structural tensor matrix of the motion area for:

[0023] (1)

[0024] in , , .

[0025] 2) Since the matrix The eigenvalues ​​in and It can represent the contrast in the direction of the eigenvector and the overall trend of grayscale changes in the window, which is represented by and Characterizing texture features of motion region frames ;

[0026] (2)

[0027] (3)

[0028] The shape feature The calculation method is:

[0029] 1) Use Harri corner point algorithm to extract the boundary corner point set of the motion area , , Indicates the length of the motion area boundary;

[0030] 2) Using boundary function Describe the boundaries of the motion area,

[0031] (4)

[0032] in Indicates the center of the motion area;

[0033] 3) Based on the boundary corner As a boundary function The independent variable is Next Linear convolution with Gaussian kernel function is obtained , then As a shape feature to characterize the motion area.

[0034] (5)

[0035] in The scale is Gaussian kernel function;

[0036] Step 4. Use the trained target BP neural network algorithm to identify the safety behavior of power system workers in the movement area.

[0037] Furthermore, the training process of the BP neural network algorithm in step 4 is:

[0038] (1) The BP neural network algorithm first segments the detected motion area based on the sliding window mechanism, taking each cell as a sample and calculating the five types of feature descriptors in each cell, that is, each sample contains 5 features;

[0039] (2) A three-layer BP neural network structure is used as the sample training structure. The three-layer structure includes: input layer, hidden layer and output layer, and bias units are added to the hidden layer and output layer respectively. At this time, the BP neural network is constructed;

[0040] (3) After the BP neural network is constructed, two steps of forward propagation and back propagation will be performed; in the forward propagation process, the sigmoid function is selected as the activation function from the input layer to the hidden layer; in the back propagation process, the cross entropy loss function is used instead of the mean square error function to update the weights and biases; the Taylor formula is introduced to determine the error accuracy;

[0041] Repeat (1) to (3) to train the target BP neural network and save it to meet the accuracy range. The set of weight ranges for all normative operations , Deviation Range Set and offset range set .

[0042] Furthermore, the safety identification process of the power system staff behavior in step 4 is as follows: the motion area extracted from the feature in step 3 is input into the target BP neural network for feature classification; the classification is specifically as follows: , Deviation Range Set and offset range set Randomly select values ​​for feature classification, and judge whether the error of the output layer result is within the specified accuracy range, so as to safely identify the behavior of power system staff.

[0043] The advantages of the present invention are:

[0044] The system provided by the present invention divides video information into monitoring video stream and personnel behavior stream through the 5G MEC edge computing server. Target recognition is only for the personnel behavior stream, and the amount of calculation is greatly reduced. Different transmission paths and adapted optical fibers are deployed according to different information flows, which can minimize the optical fiber cost and data processing volume, and achieve data and bandwidth matching, which can not only ensure the timeliness of transmission of different types of data streams, but also save equipment deployment costs.

[0045] The present invention can solve the problems of transmission delay in traditional power systems based on 4G network video collection, and the collected video being affected by light, electromagnetic interference, etc., resulting in low quality. At the same time, it can identify and warn the behavior of power system personnel, reducing the safety hazards of stable operation of the power system. The shunt design structure can further improve the transmission rate of the 5G network, and shunt the 5G signal to realize automatic identification of personnel behavior.

[0046] The present invention is based on the existing Itti algorithm. Through improvement, it realizes the extraction of a total of five types of features, including texture features and shape features in addition to conventional brightness features, color features, and direction features, and provides a texture feature and shape feature extraction process. The present invention introduces texture features to solve the defect that the original algorithm does not clearly characterize the texture area, and the introduction of shape features will more accurately locate the boundaries of different motion areas. The present invention selects five types of features extracted in advance as sub-features for constructing a BP neural network. After repeated training of the correct behavior of power system personnel based on the neural network, a neural network architecture for the behavior of power system personnel will be obtained, which can more accurately and reliably identify the unsafe behavior of power system personnel. At the same time, compared with the traditional neural network architecture, the neural network architecture based on the 5G network has higher computing efficiency and better recognition effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a structural schematic diagram of the automatic identification system of personnel behavior in the power system with 5G MEC offload in the implementation of the present invention;

[0048] Figure 2 This is a schematic diagram of the networking structure of the automatic identification system for power system personnel behavior with 5G MEC offload in an embodiment of the present invention;

[0049] Figure 3 This is a block diagram of the MEC edge computing server diversion structure in an embodiment of the present application;

[0050] Figure 4 This is a block diagram of the overall structure of the power system personnel safety behavior identification system based on artificial intelligence and 5G in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] Glossary:

[0053] AMF: access and mobility management; SMF: session management; UPF: user plane function; PCF: trusted domain; NEF: untrusted domain; MEC AF: edge computing service function; AF Request: service function request; UPF (UL CL): user plane function (uplink offload); UE: terminal; RAN: access network; UPF anchor 1: anchor point 1 UPF; UPF anchor 2: anchor point 2 UPF; N1, N2, N3, N4, N6, N9: represent different interfaces.

[0054] This embodiment provides a 5G MEC-divided power system personnel behavior automatic identification system. According to the deployment policy of "safety partitioning, network dedicated, horizontal isolation, and vertical encryption" of the power system, the deployment location and layout relationship of this system are described in detail as follows:

[0055] 1. First, in the safety zone I (control Figure 1 In the on-site area), 4K variable-focus HD cameras are deployed on-site. The number of cameras deployed will change accordingly according to the different on-site locations. This type of camera has built-in anti-light sensitive elements and anti-magnetic interference elements, which can be immune to the light and electromagnetic interference of the surrounding environment; wireless CPE is deployed near the 4K variable-focus HD camera to receive the signal of the 4K variable-focus HD camera. In order to ensure the reliability and stability of the signal transmission from the 4K variable-focus wide-angle HD camera to the CPE, the best CPE deployment position between every two 4K variable-focus HD cameras will be determined based on the principle of mirror reflection; considering the upper limit of the transmission distance of the CPE equipment, the deployment position of the 5G base station will be determined within the distance range of 3-5km to ensure the stability of the transmission signal of the entire system.

[0056] The use of wireless transmission in Zone I can save the cost of laying optical fiber and is highly economical. At the same time, wireless transmission can save space and site requirements.

[0057] 2. In the safety zone II (control Figure 1In the communication room area in the center), the isolation network gateway is connected to the 5G base station through type A optical fiber. The isolation network gateway is deployed in the communication room. Considering the cost of laying optical fiber, the communication room closest to the 5G base station will be selected to place the equipment to ensure the economy of system construction. At the same time, this type of optical fiber can ensure the maximum throughput of the transmission signal and the stability of the system signal transmission; the isolation network gateway will be connected to the Winut core switch through type A optical fiber. The core switch is deployed in the communication room. In order to improve the security of the system, a Winut bastion machine will be deployed in bypass of the Winut core switch to provide abnormal detection, authorized access and audit functions for the transmission signal; the Winut core switch will be connected to the MEC edge computing server through type A optical fiber. The MEC edge computing server is deployed in the communication room. room; after the MEC edge computing server diverts the signal, the real-time monitoring video stream on site is transmitted to the operator station host based on the A-type optical fiber through the Winut firewall. The Winut firewall is deployed in the communication room, and the operator station host is located in the control room, which is convenient for centralized control of the transmission results; the MEC edge computing server transmits the personnel behavior video stream based on the B-type optical fiber through the Winut firewall to the algorithm server for calculation and identification. The bandwidth of the B-type optical fiber and the bandwidth of the A-type optical fiber will be selected according to the uplink limit of the amount of transmitted data to ensure the optimization of bandwidth usage and save bandwidth resources. At the same time, the filtering of the Winut firewall will improve the security of the system; the algorithm server will be connected to the operator station host through the B-type optical fiber, and the identification results will be fed back to the operator station host in real time for monitoring and alarm. The equipment connection in Zone II adopts dual-line redundant deployment to improve the robustness of the system. In this embodiment, the A-type optical fiber model can be G.657.A2 or G.654.D; the B-type optical fiber model can be G.652.D or G.652.

[0058] In this embodiment, the 4K variable-focus high-definition camera has built-in anti-light sensitive elements and anti-magnetic interference elements to improve the camera effect.

[0059] The specific process of video signal splitting in this embodiment is as follows:

[0060] The wireless CPE is used to receive the video signal transmitted by the 4K variable focal length high-definition camera, and forward the signal to the small 5G base station for transmission through the 5G network. The MEC edge computing server will perform the first diversion of the signal transmitted by the small 5G base station, so that the 5G signal avoids the operator's core network and is transmitted within the power system plant. Then, the video stream and the personnel behavior stream are diverted for the second time through the diversion design structure based on the first diversion result, and the diversion result is transmitted to the firewall for filtering to realize automatic identification of personnel behavior; the algorithm server will perform moving target detection on the personnel behavior stream transmitted by the MEC edge computing server, identify the personnel movement area in the video frame, and extract features of the movement area, and use the extracted features as samples for training and modeling in the neural network algorithm, and finally complete the classification and identification of personnel movement behavior; the operator station host will receive the personnel behavior results identified by the algorithm server, and warn of unsafe behavior, and receive the real-time on-site images transmitted by the MEC edge computing server for monitoring.

[0061] In this embodiment, the MEC edge computing server is configured with a diversion structure for implementing the second diversion, specifically:

[0062] The terminal UE is connected in communication with MEC AF, MEC AF is connected in communication with NEF, NEF is connected in communication with PCF, NEF and PCF are connected in communication with AMF and SMF respectively, AMF is connected in communication with UE and RAN respectively, UE is connected in communication with RAN, SMF and RAN are connected in communication with UPF (UL-CL) respectively. UPF (UL-CL) is set with two anchor points UPF anchor1 and UPF anchor2 to connect the monitoring video stream and personnel behavior stream to two different paths of the same data network. The specific execution process is as follows: the MEC edge computing server interacts with the 5G network. At this time, the UE control layer service initiates an edge service request to the MEC edge computing server AF, that is, the first diversion service initiates an edge service request to the MEC edge computing server. At this time, AF sends an AF edge service request to NEF or PCF, including a series of parameter information such as application ID, application location, UE information, space and time validity conditions, etc. PCF or NEF generates corresponding rules based on the parameter information and its own strategy, and SMF selects the best UPF for the current service. At this time, the ULCL uplink diversion strategy in UPF is selected to divert the on-site real-time monitoring video stream and the personnel behavior video stream. UL CL uses the method of setting anchor points to connect the on-site real-time monitoring video stream and the personnel behavior video stream to two different paths of the same data network to achieve diversion, and configure the target service to be transmitted to the firewall through the N6 interface, and finally achieve the diversion of the target service and the automatic identification of personnel behavior after diversion. Among them, the PCF's own strategy is a strategy that is well known to 5G technology itself, and PCF itself is a policy control; the generated rules are PCF itself generating PCC rules in combination with parameter information, and its specific generation process is 5G existing technology and is not within the scope of consideration of the present invention; through the above process, the session management SMF can select the best UPF; the whole process does not involve new technologies.

[0063] Reference Figure 4 After the MEC server transmits the personnel behavior flow to the artificial intelligence algorithm server, the artificial intelligence algorithm server will complete the motion area detection, feature extraction and recognition of the power system personnel through the built-in algorithm. The specific execution steps are as follows:

[0064] VIBE algorithm detects motion areas in video frames. The video stream transmitted from the MEC server to the artificial intelligence algorithm server first needs to undergo preprocessing of the video frames with the VIBE algorithm for motion target detection. This process includes background model initialization, foreground detection, and background model update of the video frames. Since the background modeling method can detect motion targets in a more real-time manner and can detect motion targets more accurately, this process will ultimately extract the motion target area of ​​the video frame, that is, the behavioral motion area of ​​the power system staff.

[0065] Improved Itti algorithm extracts motion region features. The motion region obtained by VIBE algorithm preprocessing will be subjected to feature extraction by improved Itti algorithm. The traditional Itti algorithm extracts brightness features, color features and direction features in the image as sub-features describing the current image. In order to more accurately describe the behavior of power production personnel from multi-scale, multi-level and multi-perspective aspects, texture features and shape features are introduced as sub-features describing the image based on the idea of ​​image pyramid. Since the original Itti algorithm only extracts the brightness features, color features and direction features of the image frame as the features of the current image frame, the final identified area is an irregular area. At the same time, the texture representation of areas with different motion frequencies such as different motion areas and static areas is not obvious. Therefore, in order to more accurately identify the personnel motion behavior area in the image frame, texture features will be introduced to solve the defect that the original algorithm does not clearly represent the texture area. The introduction of shape features will more accurately locate the boundaries of different motion areas.

[0066] The texture features The calculation method is:

[0067] 1) For sports areas ,make and Represent the horizontal and vertical gradients of the motion area respectively, then the structural tensor matrix of the motion area for:

[0068] (1)

[0069] in , , .

[0070] 2) Since the matrix The eigenvalues ​​in and It can represent the contrast in the direction of the eigenvector and the overall trend of grayscale changes in the window, which is represented by and Characterizing texture features of motion region frames .

[0071] (2)

[0072] (3)

[0073] The shape feature The calculation method is:

[0074] 1) Use Harri corner point algorithm to extract the boundary corner point set of the motion area , , Indicates the length of the motion region boundary.

[0075] 2) Using boundary function Describe the boundaries of the motion area,

[0076] (4)

[0077] in Indicates the center of the motion area.

[0078] 3) Based on the boundary corner As a boundary function The independent variable is Next Linear convolution with Gaussian kernel function is obtained , then As a shape feature to characterize the motion area.

[0079] (5)

[0080] in The scale is Gaussian kernel function.

[0081] Improved BP neural network algorithm for motion area recognition algorithm First, the detected motion area is segmented based on the sliding window mechanism, each cell is used as a sample, and the five types of feature descriptors (brightness feature, color feature, direction feature, texture feature and shape feature) in each cell are calculated, that is, each sample contains 5 features, and a three-layer BP neural network structure is used as the sample training structure. The three-layer structure includes: input layer, hidden layer and output layer, and bias units are attached to the hidden layer and the output layer respectively. At this time, the BP neural network is constructed. After the BP neural network is constructed, two steps of forward propagation and back propagation will be performed. In the forward propagation process, the sigmoid function is selected as the activation function from the input layer to the hidden layer. Considering that the Softmax function can classify samples more accurately, the Softmax function is used as the activation function from the hidden layer to the output layer here; in the back propagation process, the traditional BP neural network algorithm updates the weights with the mean square error loss function ,deviation and bias , which converges slowly and easily falls into local minima, so the cross entropy loss function is used instead of the mean square error function to update the weights and biases. Considering the accuracy of the error of the termination iteration condition This will affect the final result of the classification. Here, the Taylor formula is introduced to more accurately determine the error accuracy. (4) Motion area detection, feature extraction and recognition

[0082] 1) First, collect the video of the power system staff demonstrating all the standard operating behaviors of the power plant, and detect the motion area of ​​the video frame according to step (1), then perform feature extraction operation on the detected motion area according to step (2), and perform step (3) on the basis of step (2), repeatedly train the newly created BP neural network, and save the results that meet the accuracy range. The set of weight ranges for all normative operations , Deviation Range Set and offset range set .

[0083] 2) The daily collected videos of power system workers working on site are also subjected to target detection and feature extraction according to steps (1) and (2), and feature classification is performed through step (3). Here, the weight range set , Deviation Range Set and offset range set Randomly select values ​​for feature classification, and judge whether the error of the output layer result is within the specified accuracy range, so as to safely identify the behavior of power system staff.

[0084] The algorithm server transmits the results of personnel behavior recognition to the industrial PC, which will issue an alarm for unsafe behavior, and the on-site staff will take the next step based on the unsafe behavior alarmed.

[0085] In summary, the application of safe behavior recognition of power system personnel based on artificial intelligence and 5G can be realized.

[0086] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A 5G MEC-based power system personnel behavior automatic identification system, characterized in that: On-site video acquisition equipment, wireless CPE, 5G base station, MEC edge computing server, algorithm server, operator station host; The on-site video acquisition device is communicatively connected to the wireless CPE, the wireless CPE is communicatively connected to the 5G base station, the 5G base station is communicatively connected to the MEC edge computing server; the MEC edge computing server is communicatively connected to the algorithm server and the operator station host respectively; the algorithm server is communicatively connected to the operator station host; The specific process of video signal diversion is as follows: The wireless CPE receives the video signal transmitted on site, and forwards the video signal to the 5G base station and transmits it to the MEC edge computing server through the 5G network, avoiding the public network and realizing the first diversion; then the MEC edge computing server performs a second diversion on the first diversion result to obtain the monitoring video stream and the personnel behavior stream, and transmits the diversion result to the firewall for filtering. The filtered monitoring video stream is sent to the host in the central control room, and the personnel behavior stream is sent to the algorithm server; the algorithm server performs moving target detection on the personnel behavior stream, identifies the personnel movement area in the video frame, and extracts features of the movement area, and uses the extracted features as samples for training and modeling in the neural network algorithm, and finally completes the classification and recognition of personnel movement behavior; The MEC edge computing server is configured with a diversion structure for implementing the second diversion, specifically: The MEC edge computing server is configured with access and mobility management unit AMF, session management unit SMF, user plane function unit UPF, trusted domain PCF, untrusted domain NEF, edge computing service function unit MEC AF, service function request unit AF Request, user plane function unit UPF (UL CL), terminal UE, access network RAN, anchor point 1 UPF anchor1, anchor point 2 UPF anchor2; The terminal UE is connected in communication with the MEC AF, the MEC AF is connected in communication with the NEF, the NEF is connected in communication with the PCF, the NEF and PCF are connected in communication with the AMF and SMF respectively, the AMF is connected in communication with the UE and the RAN respectively, the UE is connected in communication with the RAN, the SMF and the RAN are connected in communication with the UPF (UL-CL) respectively, and the UPF (UL-CL) is provided with two anchor points UPF anchor1 and UPF anchor2 to connect the monitoring video stream and the personnel behavior stream to two different paths of the same data network; Front-end video acquisition equipment, wireless CPE, and 5G base stations are deployed on site; there are multiple front-end video acquisition devices, and a wireless CPE is deployed between two adjacent front-end video acquisition devices; Deploy 5G base stations within 3-5m of wireless CPE; The MEC edge computing server and algorithm server are arranged in the communication room; An isolation network gate is also arranged in the communication room; the 5G base station is connected to the isolation network gate through A-type optical fiber; the isolation network gate is connected to the switch in communication, and the switch is connected to the MEC edge computing server; a bastion host is deployed in the bypass of the switch; after the MEC edge computing server diverts the signal, the real-time monitoring video stream on site is transmitted to the operator station host through the Winut firewall based on the A-type optical fiber, and the operator station host is located in the control room; The MEC edge computing server transmits the personnel behavior video stream to the algorithm server for calculation and identification based on the B-type optical fiber through the Winut firewall. The bandwidth of the B-type optical fiber and the bandwidth of the A-type optical fiber will be selected according to the uplink limit of the amount of transmitted data; The algorithm server will be connected to the operator station host via a B-type optical fiber.

2. According to the 5G MEC offloaded power system personnel behavior automatic identification system according to claim 1, it is characterized in that: The algorithm server is configured with a power system personnel safety behavior recognition program, and the specific recognition process is as follows: Step 1. Obtain video information of power workers' operation behavior; Step 2. Use the VIBE algorithm to detect the motion area in the video information and obtain the motion area; Step 3. Use the Itti algorithm to extract motion area features, which include five types of features: brightness feature, color feature, direction feature, texture feature, and shape feature; Step 4. Use the trained target BP neural network algorithm to safely identify the behavior of power system workers in the movement area.

3. According to claim 2, a 5G MEC offloaded power system personnel behavior automatic identification system is characterized in that: The calculation method of the texture feature D in step 3 is: 1) For the motion region L(x,y), let L x and L y Represent the horizontal and vertical gradients of the motion area respectively, then the structure tensor matrix T of the motion area is: in 2) Since the eigenvalues ​​λ1 and λ2 in the matrix T can represent the contrast in the direction of the eigenvector and the overall trend of grayscale changes in the window, the texture features D of the motion region frame are represented by λ1 and λ2; 4. According to claim 3, a 5G MEC offloaded power system personnel behavior automatic identification system is characterized in that: The calculation method of the shape feature δ(σ,i) is: 1) Harri corner point algorithm is used to extract the boundary corner point set x(i) of the motion area, i = 1, 2, ... M, M represents the length of the motion area boundary; 2) Use boundary function δ(i) to describe the boundary of the motion area, δ(i)=[x(i)-C a (i)-C b (i)] (4) Among them (C a ,C b ) represents the center of the motion area; 3) Take the boundary corner point x(i) as the independent variable of the boundary function δ(i), and perform linear convolution of δ(i) with the Gaussian kernel function under scale σ to obtain δ(σ,i), and then use δ(σ,i) as the shape feature to characterize the motion area; δ(σ,i)=[x(i)*g(σ),y(i)*g(σ)] (5) Where g(σ) represents the Gaussian kernel function with scale σ.

5. According to claim 2, a 5G MEC offloaded power system personnel behavior automatic identification system is characterized in that: The training process of the BP neural network algorithm in step 4 is: (1) The BP neural network algorithm first segments the detected motion area based on the sliding window mechanism, taking each cell as a sample and calculating the five types of feature descriptors in each cell, that is, each sample contains 5 features; (2) A three-layer BP neural network structure is used as the sample training structure. The three-layer structure includes: input layer, hidden layer and output layer, and bias units are added to the hidden layer and output layer respectively. At this time, the BP neural network is constructed; (3) After the BP neural network is constructed, two steps of forward propagation and back propagation will be carried out; in the forward propagation process, the sigmoid function is selected as the activation function from the input layer to the hidden layer; in the back propagation process, the cross entropy loss function is used instead of the mean square error function to update the weights and biases; the Taylor formula is introduced to determine the error accuracy; Repeat (1) to (3) to train the target BP neural network and save the neural network that meets the accuracy range. The set of weight ranges for all normative operations Deviation range set and offset range set 6. According to claim 5, a 5G MEC offloaded power system personnel behavior automatic identification system is characterized in that: The safety identification process of the power system staff behavior in step 4 is as follows: the motion area extracted from the feature in step 3 is input into the target BP neural network for feature classification; the classification is specifically as follows: Deviation range set and offset range set Randomly select values ​​for feature classification, and judge whether the error of the output layer result is within the specified accuracy range, so as to safely identify the behavior of power system staff.

Images