Substation risk early warning method and system based on edge AI analyzer

Abstract

The application relates to the technical field of risk early warning, and particularly discloses a transformer substation risk early warning method and system based on an edge AI analyzer, which comprises the following steps: determining the position of the edge AI analyzer and the application range thereof according to the determined position of a gun; receiving the video obtained by the gun in the edge AI analyzer at a fixed time, performing target positioning on the video, determining a target frame, identifying the content in the target frame based on a preset convolution feature table, and outputting a risk type; and when the output of the identification of the content in the target frame based on the preset convolution feature table is empty, identifying the content in the target frame based on an AI identification model in the edge AI analyzer, and outputting a risk type. The local identification module based on the convolution kernel table and the AI identification module based on the AI model are built in the edge AI analyzer, the transformer substation risk is identified in a gradient mode, the identification speed is relatively high, and the identification range is relatively large.

Description

Technical Field

[0001] This invention relates to the field of risk warning technology, specifically a substation risk warning method and system based on an edge AI analyzer. Background Technology

[0002] The Visual AI Edge Analyzer System – An edge analyzer is an AI model computing device with visual AI model inference capabilities. It processes video streams by extracting frames, inputting continuous images into a convolutional neural network for image recognition, and combining this with business logic programs for logical judgment, ultimately generating alarm events.

[0003] Visual AI edge analyzers have many applications, including mining, coal, oil, petrochemical, and power industries. This application focuses on the power sector. Existing power risk warning processes rely heavily on manual labor, as illustrated in patent CN112734254B. Risk warnings depend on manual processes, specifically determining the target construction procedure from a pre-set 3D construction model, obtaining risk point information corresponding to the target procedure, generating risk warning information based on the risk point information, and sending the risk warning information to the terminal device. Although an intelligent construction procedure determination scheme is introduced in this process, the level of intelligence is not high, and the work of risk monitoring personnel is extremely tedious. In today's rapidly developing AI technology landscape, how to intelligently optimize inspection tasks using AI technology is the technical problem this invention aims to solve. Summary of the Invention

[0004] The purpose of this invention is to provide a substation risk early warning method and system based on an edge AI analyzer to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A substation risk early warning method based on an edge AI analyzer, the method comprising:

[0007] Obtain the equipment distribution information of the substation, and determine the location and parameters of the gun based on the equipment distribution information;

[0008] The location of the edge AI analyzer and its application range are determined based on the identified gun position; the application range includes the identification of the gun to which the edge AI analyzer is connected;

[0009] The edge AI analyzer periodically receives video footage acquired by the camera, performs target localization on the video, determines the target bounding box, identifies the content within the target bounding box based on a preset convolutional feature table, and outputs the risk type; the convolutional feature table includes preset risk types and their convolutional kernels;

[0010] When the output of the identification of the content in the target box based on the preset convolutional feature table is empty, the AI ​​recognition model in the edge AI analyzer identifies the content in the target box and outputs the risk type.

[0011] The risk types of each camera are statistically analyzed, and the application scope of the edge AI analyzer and the video reception cycle of each camera are updated.

[0012] As a further aspect of the present invention: the step of obtaining equipment distribution information of the substation and determining the location and parameters of the gun based on the equipment distribution information includes:

[0013] Obtain the equipment registration form of the substation and query the equipment level and location of each piece of equipment;

[0014] The scope of influence and the values ​​of each point within the scope of influence are determined based on the equipment level.

[0015] Centered on the device location, calculate the range of influence and the values ​​of each point within the range of influence, and create a monitoring area containing numerical values;

[0016] By overlaying the monitoring areas of all devices, the values ​​at each point in the substation area are obtained;

[0017] The substation area is divided into sub-regions based on the inscribed rectangle of the camera's monitoring range;

[0018] The values ​​at each point in the sub-region are accumulated to obtain a numerical sum, and the bullet sharpness is determined based on the numerical sum.

[0019] The values ​​for each point within the affected area are:

[0020] In the formula, N is the center value determined by the equipment level, R is the range radius determined by the equipment level, I(ρ) represents the value of the point at radius ρ, and α is a preset correction coefficient.

[0021] As a further aspect of the present invention: the step of determining the location and application range of the edge AI analyzer based on the determined gun position includes:

[0022] The resource allocation of the bolt is determined based on the bolt's clarity;

[0023] Cluster the guns based on the resource proportions;

[0024] Calculate the cluster center for each type of gun, and select the cluster center as the location of the edge AI analyzer;

[0025] Obtain the tags of the corresponding classes of edge AI analyzers to define the application scope of the edge AI analyzers.

[0026] As a further aspect of the present invention: the steps of receiving video acquired by the camera at regular intervals based on the edge AI analyzer, locating targets in the video, determining target bounding boxes, identifying the content within the target bounding boxes based on a preset convolutional feature table, and outputting the risk type include:

[0027] The edge AI analyzer periodically receives the video acquired by the camera and converts the video into an image sequence;

[0028] By successively subtracting adjacent images, a difference image is obtained;

[0029] Traverse the pixels in the difference image, mark the pixels with pixel values ​​greater than a preset threshold, connect adjacent marked pixels to obtain the target region;

[0030] Remove target regions with an area smaller than a preset area threshold to obtain the target bounding box;

[0031] Based on a pre-defined convolutional feature table, the content within the target bounding box is identified to obtain the risk type of each image.

[0032] As a further aspect of the present invention: when the output of recognizing the content in the target box based on the preset convolutional feature table is empty, the step of recognizing the content in the target box based on the AI ​​recognition model in the edge AI analyzer and outputting the risk type includes:

[0033] When the output of recognizing the content in the target box based on the preset convolutional feature table is empty, the image containing the target box is input into the AI ​​recognition model in the edge AI analyzer.

[0034] Read the output of the AI ​​recognition model and use it as the risk type of the target box.

[0035] As a further aspect of the present invention: the steps of statistically analyzing the risk type of each camera, updating the application scope of the edge AI analyzer, and the video reception cycle of each camera include:

[0036] Risk type of the gun based on image sequence reading;

[0037] Query the risk value corresponding to the risk type in the preset numerical table, and adjust the video reception cycle of the camera according to the risk value;

[0038] The correction rate is determined based on the risk value and the preset risk value threshold. The resource allocation of the gun is adjusted based on the correction rate, and the application scope of the edge AI analyzer is updated.

[0039] The correction rate is determined as follows:

[0040] P = A(F - F0); where P is the correction rate, F is the risk value, F0 is the risk value threshold, and A is a preset constant.

[0041] The present invention also provides a system for applying the above-mentioned substation risk early warning method based on edge AI analyzer, the system comprising:

[0042] The gun mounting module is used to acquire equipment distribution information of the substation and determine the gun position and its parameters based on the equipment distribution information.

[0043] The analyzer calibration module is used to determine the location and application range of the edge AI analyzer based on the determined gun position; the application range includes the identification of the gun connected to the edge AI analyzer;

[0044] The target localization module is used to periodically receive video acquired by the camera based on the edge AI analyzer, perform target localization on the video, determine the target box, identify the content in the target box based on a preset convolutional feature table, and output the risk type; the convolutional feature table includes preset risk types and their convolutional kernels;

[0045] The content recognition module is used to recognize the content in the target box based on the AI ​​recognition model in the edge AI analyzer and output the risk type when the output of the recognition of the content in the target box based on the preset convolutional feature table is empty.

[0046] The process update module is used to statistically analyze the risk type of each camera, update the application scope of the edge AI analyzer, and update the video reception cycle of each camera.

[0047] As a further aspect of the present invention: the bolt mounting module includes:

[0048] The equipment query unit is used to obtain the equipment registration form of the substation and query the equipment level and location of each piece of equipment.

[0049] A numerical determination unit is used to determine the scope of influence and the value of each point within the scope of influence based on the equipment level.

[0050] The monitoring area determination unit is used to create a monitoring area containing numerical values ​​by taking the device location as the center, calculating the range of influence and the values ​​of each point within the range of influence;

[0051] The monitoring area overlay unit is used to overlay the monitoring areas of all devices to obtain the values ​​at each point in the substation area.

[0052] The area segmentation unit is used to divide the substation area into sub-regions based on the inscribed rectangle of the monitoring range of the camera.

[0053] The numerical accumulation unit is used to accumulate the values ​​at each point in the sub-region to obtain a numerical sum, and to determine the gun's sharpness based on the numerical sum.

[0054] The values ​​for each point within the affected area are:

[0055] In the formula, N is the center value determined by the equipment level, R is the range radius determined by the equipment level, I(ρ) represents the value of the point at radius ρ, and α is a preset correction coefficient.

[0056] As a further aspect of the present invention: the analyzer calibration module includes:

[0057] The resource allocation determination unit is used to determine the resource allocation of the bolt based on the bolt's clarity.

[0058] A gun clustering unit is used to cluster guns according to the resource proportion;

[0059] The location setting unit is used to calculate the cluster center of each type of gun and select the cluster center as the location of the edge AI analyzer.

[0060] The tag query unit is used to obtain the tags of the corresponding class of the edge AI analyzer, which serves as the application scope of the edge AI analyzer.

[0061] As a further aspect of the present invention: the target positioning module includes:

[0062] The image sequence generation unit is used to periodically receive the video acquired by the camera based on the edge AI analyzer and convert the video into an image sequence.

[0063] The image subtraction unit is used to successively subtract adjacent images to obtain a difference image;

[0064] The pixel connection unit is used to traverse the pixels in the difference image, mark the pixels with pixel values ​​greater than a preset threshold, and connect adjacent marked pixels to obtain the target region.

[0065] The region elimination unit is used to eliminate target regions with an area smaller than a preset area threshold to obtain the target bounding box;

[0066] The content recognition unit is used to recognize the content in the target box based on a preset convolutional feature table to obtain the risk type of each image.

[0067] Compared with the prior art, the beneficial effects of the present invention are: the present invention determines the distribution information of the guns based on the power equipment, installs an edge AI analyzer based on the distribution information of the guns, and integrates a local recognition module based on a convolution kernel table and an AI recognition module based on an AI model into the edge AI analyzer, thereby identifying substation risks in a gradient manner, with faster recognition speed and wider recognition scope. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention.

[0069] Figure 1 This is a flowchart of a substation risk early warning method based on an edge AI analyzer.

[0070] Figure 2 This is a block diagram of the composition structure of a substation risk early warning system based on an edge AI analyzer. Detailed Implementation

[0071] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0072] Figure 1 This is a flowchart of a substation risk early warning method based on an edge AI analyzer. In this embodiment of the invention, a substation risk early warning method based on an edge AI analyzer includes:

[0073] Step S100: Obtain the equipment distribution information of the substation, and determine the location and parameters of the gun based on the equipment distribution information;

[0074] A substation is a place in a power system that transforms voltage and current, receives electrical energy, and distributes electrical energy. It contains various electrical equipment. The location and parameters of each electrical device constitute the equipment distribution information. By obtaining the distribution information of the electrical equipment, the installation location and installation parameters of the camera can be determined. The camera can be understood as a fixed camera with high precision and a fixed shooting frame. Accordingly, the camera parameters are the resolution of the camera.

[0075] Step S200: Determine the location of the edge AI analyzer and its application range based on the determined gun position; the application range includes the identifier of the gun connected to the edge AI analyzer;

[0076] There is a data interaction channel between the edge AI analyzer and the receiver. The location of the edge AI analyzer can be determined based on the location of the receiver. The receiver corresponding to each edge AI analyzer is the application range of the edge AI analyzer. The application range can be composed of the identifier of the receiver connected to the edge AI analyzer, which is generally a number.

[0077] Step S300: Based on the edge AI analyzer, the video acquired by the camera is received periodically, the target is located in the video, the target box is determined, the content in the target box is identified based on the preset convolutional feature table, and the risk type is output; the convolutional feature table includes the preset risk type and its convolutional kernel;

[0078] For each edge AI analyzer, it interacts with a portion of the cameras, periodically receives videos acquired by the cameras, performs target localization on each video to obtain target bounding boxes, and first identifies the content within the target bounding boxes based on a preset convolutional feature table to determine the risk type. This process is essentially an image traversal and comparison process and does not involve the AI ​​recognition process. The convolutional kernel table is a preset table, which is obtained by staff in advance by acquiring images under different risk types, recognizing the images, and then extracting image features, which are called convolutional kernels.

[0079] Step S400: When the output of recognizing the content in the target box based on the preset convolutional feature table is empty, the content in the target box is recognized based on the AI ​​recognition model in the edge AI analyzer, and the risk type is output.

[0080] The target recognition process based on convolutional feature tables can be performed locally. In existing technologies, the recognition speed may not differ much, but the amount of resources consumed is less. The disadvantage of the target recognition process based on convolutional feature tables is that it cannot identify risk types that have not been statistically analyzed in advance. At this time, the introduction of AI recognition models can identify unknown targets, which greatly improves the comprehensiveness of recognition.

[0081] Step S500: Analyze the risk type of each camera, update the application scope of the edge AI analyzer and the video reception cycle of each camera;

[0082] After identifying the risk type, the substation risk warning process has been completed. Based on this, considering the workload of each edge AI analyzer, this application introduces an update process. For a certain edge AI analyzer, the risk types corresponding to each camera are counted. The more risk types there are, the shorter the video reception cycle of the corresponding camera needs to be, and the more videos are acquired. Correspondingly, the number of cameras connected to the edge AI analyzer needs to be reduced.

[0083] As a preferred embodiment of the technical solution of the present invention, the step of obtaining the equipment distribution information of the substation and determining the location and parameters of the gun based on the equipment distribution information includes:

[0084] Obtain the equipment registration form of the substation and query the equipment level and location of each piece of equipment;

[0085] The scope of influence and the values ​​of each point within the scope of influence are determined based on the equipment level.

[0086] Centered on the device location, calculate the range of influence and the values ​​of each point within the range of influence, and create a monitoring area containing numerical values;

[0087] By overlaying the monitoring areas of all devices, the values ​​at each point in the substation area are obtained;

[0088] The substation area is divided into sub-regions based on the inscribed rectangle of the camera's monitoring range;

[0089] The values ​​at each point in the sub-region are accumulated to obtain a numerical sum, and the bullet sharpness is determined based on the numerical sum.

[0090] The values ​​for each point within the affected area are:

[0091] In the formula, N is the center value determined by the equipment level, R is the range radius determined by the equipment level, I(ρ) represents the value of the point at radius ρ, and α is a preset correction coefficient.

[0092] Furthermore, the values ​​at each point can be obtained using a two-dimensional Gaussian distribution model, resulting in even greater smoothness.

[0093] In one example of the technical solution of this invention, the power equipment in the substation is stored in an equipment registration form, which includes the equipment location and equipment level of each power equipment. The equipment level represents the importance of the power equipment. The scope of influence is determined according to the equipment level. The higher the equipment level, the larger the scope of influence. Combined with the equipment location, an area of ​​influence can be determined, called the monitoring area. The monitoring area is generally circular.

[0094] By overlaying the monitoring areas of all devices, the values ​​at each point in the substation area are obtained, and the values ​​at each point reflect the importance of the corresponding point. Based on this, the monitoring range of the camera is obtained. This application requires global monitoring of the entire substation, and the set of monitoring ranges of all cameras must be larger than the substation area. Therefore, the entire area of ​​the substation can be divided by the monitoring range of the cameras to obtain sub-areas.

[0095] Since each point has been assigned a value indicating its importance, the sum of the values ​​at each point in the sub-region can be obtained. The gun's sharpness can be determined based on the sum of the values. The gun's sharpness is directly proportional to the sum of the values; that is, the more important the area corresponding to the gun, the higher the sharpness should be.

[0096] As a preferred embodiment of the technical solution of the present invention, the step of determining the position and application range of the edge AI analyzer based on the determined gun position includes:

[0097] The resource allocation of the bolt is determined based on the bolt's clarity;

[0098] Cluster the guns based on the resource proportions;

[0099] Calculate the cluster center for each type of gun, and select the cluster center as the location of the edge AI analyzer;

[0100] Obtain the tags of the corresponding classes of edge AI analyzers to define the application scope of the edge AI analyzers.

[0101] The higher the resolution of the video, the greater the resource requirements and the higher the resource ratio for processing the corresponding video. By clustering video cameras based on their resource ratio, multiple categories of video cameras can be obtained. For a category of video cameras with a higher resource ratio, the selected edge AI analyzer should have higher performance, and vice versa. In order to minimize the transmission pressure during the data transmission process (between the video camera and the edge AI analyzer), the cluster center is selected as the location of the edge AI analyzer. The labels of the video cameras corresponding to the edge AI analyzer are obtained, which serve as the application scope of the edge AI analyzer.

[0102] As a preferred embodiment of the technical solution of the present invention, the steps of receiving video acquired by the camera at regular intervals based on the edge AI analyzer, locating targets in the video, determining target boxes, identifying the content in the target boxes based on a preset convolutional feature table, and outputting the risk type include:

[0103] The edge AI analyzer periodically receives the video acquired by the camera and converts the video into an image sequence;

[0104] By successively subtracting adjacent images, a difference image is obtained;

[0105] Traverse the pixels in the difference image, mark the pixels with pixel values ​​greater than a preset threshold, connect adjacent marked pixels to obtain the target region;

[0106] Remove target regions with an area smaller than a preset area threshold to obtain the target bounding box;

[0107] Based on a pre-defined convolutional feature table, the content within the target bounding box is identified to obtain the risk type of each image.

[0108] In one example of the technical solution of this invention, the target localization process and the application process of the convolutional feature table are described in detail. Based on the edge AI analyzer, the video acquired by the camera is received periodically and converted into an image sequence (audio is removed). The difference between adjacent images is calculated to obtain a difference image. For surveillance video, only the pixels corresponding to moving objects will have differences; other pixels will be close to zero after the difference is calculated. The pixels in the difference image are traversed, and pixels with pixel values ​​greater than a preset threshold are marked. Adjacent marked pixels are connected to obtain the target region, which is the active area where the moving object is located. Target regions with particularly small areas (noise) are removed, and the remaining target regions are taken as the regions of the real moving objects. The matrix shape of the region is obtained, called the target box. The content in the target box is identified based on the preset convolutional feature table. After a successful match, the risk type corresponding to the target box is obtained.

[0109] As a preferred embodiment of the technical solution of the present invention, the step of identifying the content in the target box based on the preset convolutional feature table and outputting the risk type when the output is empty includes:

[0110] When the output of recognizing the content in the target box based on the preset convolutional feature table is empty, the image containing the target box is input into the AI ​​recognition model in the edge AI analyzer.

[0111] Read the output of the AI ​​recognition model and use it as the risk type of the target box.

[0112] When the output of the recognition of the content in the target box based on the preset convolutional feature table is empty, that is, when the recognition of the content in the target box based on the preset convolutional feature table fails, the AI ​​recognition model in the edge AI analyzer is applied to recognize the target. Based on the already strong AI capability, the recognition breadth and recognition depth of this application are greatly improved.

[0113] It is worth mentioning that in practical applications, since there are only a limited number of risk types, convolutional feature tables are sufficient to cover most types, and the application frequency of AI recognition models is actually very low.

[0114] As a preferred embodiment of the technical solution of the present invention, the steps of statistically analyzing the risk type of each camera, updating the application scope of the edge AI analyzer, and the video reception cycle of each camera include:

[0115] Risk type of the gun based on image sequence reading;

[0116] Query the risk value corresponding to the risk type in the preset numerical table, and adjust the video reception cycle of the camera according to the risk value;

[0117] The correction rate is determined based on the risk value and the preset risk value threshold. The resource allocation of the gun is adjusted based on the correction rate, and the application scope of the edge AI analyzer is updated.

[0118] The correction rate is determined as follows:

[0119] P = A(F - F0); where P is the correction rate, F is the risk value, F0 is the risk value threshold, and A is a preset constant.

[0120] In one embodiment of the technical solution of this invention, a negative feedback adjustment scheme is provided. Based on the sequential reading of the risk type of the camera, a pre-defined numerical value representing the degree of risk (i.e., risk value) is assigned to each risk type. The risk value corresponding to the risk type is queried, and the video reception cycle of the camera is adjusted according to the risk value. The video reception cycle is inversely proportional to the risk value; the higher the risk value, the shorter the video reception cycle. Furthermore, a correction rate is determined based on the risk value to adjust the resource allocation of the camera. This affects the clustering process of the camera, i.e., which edge AI analyzers manage different cameras. Different edge AI analyzers have different performance; the higher the risk value, the higher the corresponding resource allocation.

[0121] Figure 2 The diagram shows the structural composition of a substation risk early warning system based on an edge AI analyzer. In this embodiment of the invention, a system 10 applied to the aforementioned substation risk early warning method based on an edge AI analyzer includes:

[0122] The gun mounting module 11 is used to acquire equipment distribution information of the substation and determine the gun position and its parameters based on the equipment distribution information.

[0123] The analyzer calibration module 12 is used to determine the location of the edge AI analyzer and its application range based on the determined gun position; the application range includes the identification of the gun connected to the edge AI analyzer;

[0124] The target localization module 13 is used to periodically receive the video acquired by the camera based on the edge AI analyzer, perform target localization on the video, determine the target box, identify the content in the target box based on the preset convolutional feature table, and output the risk type; the convolutional feature table includes the preset risk type and its convolutional kernel;

[0125] The content recognition module 14 is used to recognize the content in the target box based on the AI ​​recognition model in the edge AI analyzer and output the risk type when the output of the content recognition based on the preset convolutional feature table is empty.

[0126] The process update module 15 is used to count the risk type of each camera, update the application scope of the edge AI analyzer, and update the video reception cycle of each camera.

[0127] Furthermore, the bolt mounting module 11 includes:

[0128] The equipment query unit is used to obtain the equipment registration form of the substation and query the equipment level and location of each piece of equipment.

[0129] A numerical determination unit is used to determine the scope of influence and the value of each point within the scope of influence based on the equipment level.

[0130] The monitoring area determination unit is used to create a monitoring area containing numerical values ​​by taking the device location as the center, calculating the range of influence and the values ​​of each point within the range of influence;

[0131] The monitoring area overlay unit is used to overlay the monitoring areas of all devices to obtain the values ​​at each point in the substation area.

[0132] The area segmentation unit is used to divide the substation area into sub-regions based on the inscribed rectangle of the monitoring range of the camera.

[0133] The numerical accumulation unit is used to accumulate the values ​​at each point in the sub-region to obtain a numerical sum, and to determine the gun's sharpness based on the numerical sum.

[0134] The values ​​for each point within the affected area are:

[0135] In the formula, N is the center value determined by the equipment level, R is the range radius determined by the equipment level, I(ρ) represents the value of the point at radius ρ, and α is a preset correction coefficient.

[0136] Specifically, the analyzer calibration module 12 includes:

[0137] The resource allocation determination unit is used to determine the resource allocation of the bolt based on the bolt's clarity.

[0138] A gun clustering unit is used to cluster guns according to the resource proportion;

[0139] The location setting unit is used to calculate the cluster center of each type of gun and select the cluster center as the location of the edge AI analyzer.

[0140] The tag query unit is used to obtain the tags of the corresponding class of the edge AI analyzer, which serves as the application scope of the edge AI analyzer.

[0141] In addition, the target positioning module 13 includes:

[0142] The image sequence generation unit is used to periodically receive the video acquired by the camera based on the edge AI analyzer and convert the video into an image sequence.

[0143] The image subtraction unit is used to successively subtract adjacent images to obtain a difference image;

[0144] The pixel connection unit is used to traverse the pixels in the difference image, mark the pixels with pixel values ​​greater than a preset threshold, and connect adjacent marked pixels to obtain the target region.

[0145] The region elimination unit is used to eliminate target regions with an area smaller than a preset area threshold to obtain the target bounding box;

[0146] The content recognition unit is used to recognize the content in the target box based on a preset convolutional feature table to obtain the risk type of each image.

[0147] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A substation risk early warning method based on an edge AI analyzer, characterized in that, The method includes: Step S100: Obtain the equipment distribution information of the substation, and determine the location and parameters of the gun based on the equipment distribution information; Step S200: Determine the location of the edge AI analyzer and its application range based on the determined gun position; the application range includes the identifier of the gun connected to the edge AI analyzer; The step of determining the location and application range of the edge AI analyzer based on the determined gun position includes: The resource allocation of the bolt is determined based on the bolt's clarity; Cluster the guns based on the resource proportions; Calculate the cluster center for each type of gun, and select the cluster center as the location of the edge AI analyzer; Obtain the tags of the corresponding classes of edge AI analyzers to define the application scope of the edge AI analyzers; Step S300: Based on the edge AI analyzer, the video acquired by the camera is received periodically, the target is located in the video, the target box is determined, the content in the target box is identified based on the preset convolutional feature table, and the risk type is output; the convolutional feature table includes the preset risk type and its convolutional kernel; Step S400: When the output of recognizing the content in the target box based on the preset convolutional feature table is empty, the content in the target box is recognized based on the AI ​​recognition model in the edge AI analyzer, and the risk type is output. Step S500: Calculate the risk type of each camera, update the application scope of the edge AI analyzer and the video reception cycle of each camera.

2. The edge AI analyzer-based substation risk early warning method according to claim 1, characterized in that, The step of obtaining equipment distribution information of the substation and determining the location and parameters of the gun based on the equipment distribution information includes: Obtain the equipment registration form of the substation and query the equipment level and location of each piece of equipment; The scope of influence and the values ​​of each point within the scope of influence are determined based on the equipment level. Centered on the device location, calculate the range of influence and the values ​​of each point within the range of influence, and create a monitoring area containing numerical values; By overlaying the monitoring areas of all devices, the values ​​at each point in the substation area are obtained; The substation area is divided into sub-regions based on the inscribed rectangle of the camera's monitoring range; The values ​​at each point in the sub-region are accumulated to obtain a numerical sum, and the bullet sharpness is determined based on the numerical sum. The values ​​for each point within the affected area are: In the formula, It is a central value determined by the equipment level. The radius of the range is determined by the equipment level. Represents radius The value of the point at that location; This is the preset correction factor. 3.The substation risk pre-warning method based on edge AI analyzer according to claim 1, wherein, The steps of receiving video footage from the camera at regular intervals using an edge AI analyzer, locating targets in the video, determining target bounding boxes, identifying the content within the target bounding boxes based on a preset convolutional feature table, and outputting the risk type include: The edge AI analyzer periodically receives the video acquired by the camera and converts the video into an image sequence; By successively subtracting adjacent images, a difference image is obtained; Traverse the pixels in the difference image, mark the pixels with pixel values ​​greater than a preset threshold, connect adjacent marked pixels to obtain the target region; Remove target regions with an area smaller than a preset area threshold to obtain the target bounding box; Based on a pre-defined convolutional feature table, the content within the target bounding box is identified to obtain the risk type of each image.

4. The edge AI analyzer-based substation risk early warning method of claim 1, wherein, When the output of recognizing the content in the target box based on the preset convolutional feature table is empty, the step of recognizing the content in the target box based on the AI ​​recognition model in the edge AI analyzer and outputting the risk type includes: When the output of recognizing the content in the target box based on the preset convolutional feature table is empty, the image containing the target box is input into the AI ​​recognition model in the edge AI analyzer. Read the output of the AI ​​recognition model and use it as the risk type of the target box.

5. The edge AI analyzer based substation risk early warning method of claim 1, wherein, The steps of statistically analyzing the risk type of each camera, updating the application scope of the edge AI analyzer, and determining the video reception cycle of each camera include: Risk type of the gun based on image sequence reading; Query the risk value corresponding to the risk type in the preset numerical table, and adjust the video reception cycle of the camera according to the risk value; The correction rate is determined based on the risk value and the preset risk value threshold. The resource allocation of the gun is adjusted based on the correction rate, and the application scope of the edge AI analyzer is updated. The correction rate is determined as follows: ; wherein, is a correction rate, is a risk value, a risk value threshold, is a predetermined constant.

6. A substation risk early warning system based on edge AI analyzer, characterized in that, The substation risk early warning method based on an edge AI analyzer, applicable to any one of claims 1-5, wherein the substation risk early warning system (10) based on an edge AI analyzer comprises: The gun mounting module (11) is used to acquire equipment distribution information of the substation and determine the gun position and its parameters based on the equipment distribution information. The analyzer calibration module (12) is used to determine the location of the edge AI analyzer and its application range based on the determined gun position; the application range includes the identification of the gun connected to the edge AI analyzer; The target localization module (13) is used to receive the video obtained by the camera at regular intervals based on the edge AI analyzer, perform target localization on the video, determine the target box, identify the content in the target box based on the preset convolutional feature table, and output the risk type; the convolutional feature table includes the preset risk type and its convolutional kernel; The content recognition module (14) is used to recognize the content in the target box based on the AI ​​recognition model in the edge AI analyzer and output the risk type when the output of the recognition of the content in the target box based on the preset convolutional feature table is empty; The process update module (15) is used to count the risk type of each camera, update the application scope of the edge AI analyzer and the video reception cycle of each camera.

7. The edge-AI analyzer based substation risk early warning system according to claim 6, characterized in that, The bolt mounting module (11) includes: The equipment query unit is used to obtain the equipment registration form of the substation and query the equipment level and location of each piece of equipment. A numerical determination unit is used to determine the scope of influence and the value of each point within the scope of influence based on the equipment level. The monitoring area determination unit is used to create a monitoring area containing numerical values ​​by taking the device location as the center, calculating the range of influence and the values ​​of each point within the range of influence; The monitoring area overlay unit is used to overlay the monitoring areas of all devices to obtain the values ​​at each point in the substation area. The area segmentation unit is used to divide the substation area into sub-regions based on the inscribed rectangle of the monitoring range of the camera. The numerical accumulation unit is used to accumulate the values ​​at each point in the sub-region to obtain a numerical sum, and to determine the gun's sharpness based on the numerical sum. The values ​​for each point within the affected area are: In the formula, It is a central value determined by the equipment level. The radius of the range is determined by the equipment level. Represents radius The value of the point at that location; This is the preset correction factor.

8. The edge AI analyzer based substation risk early warning system according to claim 6, characterized in that, The analyzer calibration module (12) includes: The resource allocation determination unit is used to determine the resource allocation of the bolt based on the bolt's clarity. A gun clustering unit is used to cluster guns according to the resource proportion; The location setting unit is used to calculate the cluster center of each type of gun and select the cluster center as the location of the edge AI analyzer. The tag query unit is used to obtain the tags of the corresponding class of the edge AI analyzer, which serves as the application scope of the edge AI analyzer.

9. The edge-AI analyzer based substation risk early warning system according to claim 6, characterized in that, The target positioning module (13) includes: The image sequence generation unit is used to periodically receive the video acquired by the camera based on the edge AI analyzer and convert the video into an image sequence. The image subtraction unit is used to successively subtract adjacent images to obtain a difference image; The pixel connection unit is used to traverse the pixels in the difference image, mark the pixels with pixel values ​​greater than a preset threshold, and connect adjacent marked pixels to obtain the target region. The region elimination unit is used to eliminate target regions with an area smaller than a preset area threshold to obtain the target bounding box; The content recognition unit is used to recognize the content in the target box based on a preset convolutional feature table to obtain the risk type of each image.

Images