A machine vision-based intelligent inspection system and method for power systems

Abstract

This invention discloses an intelligent inspection system and method for power systems based on machine vision, relating to the field of machine vision technology. The invention simulates flight by setting an inspection path for a drone, adjusts the drone's inspection path based on the simulation results, and then allows the drone to inspect the power system scenario by adjusting the path, collecting various real-time status data and historical status data of the power system scenario. This historical status data is divided into historical image data, and multiple feature regions are marked on the historical image data. The historical image data is then spatiotemporally aligned based on the distribution of these feature regions, and a dynamic visualization scene model is constructed based on the historical image data. A real-time local multi-state device model is established based on the real-time status data. Both the real-time local multi-state device model and the dynamic visualization scene model are divided into model voxels, and anomaly locations are located based on the model voxel comparison results.

Description

Technical Field

[0001] This invention relates to the field of machine vision technology, and more specifically to an intelligent inspection system and method for power systems based on machine vision. Background Technology

[0002] As a crucial foundation for modern society, the stable and reliable operation of the power system is essential for ensuring socio-economic development and the quality of life for the people. Traditional power system inspections primarily rely on manual methods. Inspectors must conduct on-site inspections in complex power facility environments, which is not only physically demanding and inefficient, but also poses significant safety risks, especially in complex terrains and hazardous environments. Furthermore, manual inspections struggle to obtain real-time and comprehensive information on the power system's operational status, potentially failing to detect and address any outstanding faults or anomalies in a timely manner, thus increasing the risk of power system failures.

[0003] With the development of technology, some inspection methods based on traditional technologies have gradually emerged, but these methods have certain limitations. For example, some inspection methods can only acquire a single type of data, making it impossible to comprehensively and accurately assess the overall operating status of the power system; others lack effective data processing and analysis tools, making it difficult to quickly and accurately identify anomalies from large amounts of data. Furthermore, traditional inspection methods are often not scientifically sound in their inspection path planning, easily leading to blind spots and preventing the timely detection of potential safety hazards. Therefore, this paper proposes an intelligent inspection system and method for power systems based on machine vision. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent inspection system and method for power systems based on machine vision, so as to solve the problems in the background art.

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

[0006] A machine vision-based intelligent inspection method for power systems includes the following steps:

[0007] Step S1: Obtain a GIS map of the power system scenario, set the drone inspection path based on the GIS map, conduct flight simulation of the drone inspection path, adjust the drone inspection path according to the flight simulation results, and then the drone will conduct inspection in the power system scenario by adjusting the drone inspection path and collect various real-time status data.

[0008] Step S2: Obtain historical state data of the power system scenario, divide the historical state data into historical image data, and mark multiple feature regions in the historical image data. Perform spatiotemporal alignment of the historical image data according to the distribution of feature regions, and then construct a dynamic visualization scenario model based on the historical image data.

[0009] Step S3: Establish a real-time local multi-state device model based on real-time status data, divide the real-time local multi-state device model and the dynamic visualization scene model into model voxels, and locate the abnormal location based on the model voxel comparison results.

[0010] Furthermore, the process of setting up drone inspection routes includes:

[0011] Set up n drones in a power system scenario, with each drone starting at a different position in the power system scenario;

[0012] The drones are equipped with visible light imagers, infrared imagers, solar-blind ultraviolet detectors, and positioning devices, and each drone is assigned a number a1, a2, a3, ..., a n n is a natural number greater than 10;

[0013] Obtain a GIS map of the power system scenario, mark the patrol target on the GIS map, and build a 3D map of the scenario based on the GIS map;

[0014] Set a unit flight speed for the drone, and based on the flight distance of the unit flight speed in a unit time, divide the scene 3D map into several scene areas based on the flight distance, and mark the altitude value of each scene area.

[0015] The scene area associated with each patrol target is taken as the target inspection area. According to the type of power equipment corresponding to each target inspection area, the no-fly radius is marked for each target inspection area.

[0016] Set a fixed data collection range and fixed shooting angle for each drone. Then, take the shortest path, maximum inspection area coverage, no-fly radius and obstacle avoidance as prerequisites. The starting position of each drone is the inspection start point and inspection end point. Generate n drone inspection paths on the scene 3D map with different inspection start points and inspection end points, but the same spatial location.

[0017] Furthermore, the process of simulating the drone inspection path includes:

[0018] Simultaneously conduct flight simulations of each UAV inspection path in the scene 3D map at a unit flight speed, and determine whether the inspection area of ​​each target in the scene 3D map is completely covered by the data collection range of the UAV based on the flight simulation results.

[0019] If it is determined that a target inspection area is not fully covered, then adjust randomly. The adjusted drone inspection path allows the flight simulation results to cover the parts of the data collection area that were not fully covered, but at the same time, there is some overlap with the original drone inspection path.

[0020] Furthermore, the real-time scenario data acquisition process for power system scenarios includes:

[0021] The inspection paths of each drone are sent to the corresponding drones according to the inspection start point and inspection end point. Then, each drone conducts flight inspections in the power system scene at the same time according to the drone inspection path, and collects real-time status data of its power equipment through visible light imager, infrared imager and solar blind ultraviolet detector.

[0022] The real-time status data includes optical video data, infrared video data, and ultraviolet video data. At the end of each unit of time, each UAV integrates the real-time video data collected by the sensors and the real-time location information to generate real-time status data.

[0023] Furthermore, the process of annotating multiple feature regions in historical image data includes:

[0024] Obtain historical status data for all power equipment within a power system scenario over several days, assuming no anomalies.

[0025] The historical status data collected by each drone is synchronized in time. The historical video data is divided into several historical image data by frame. Then, based on the historical location information and shooting angle of each drone, the historical image data corresponding to each drone is mapped onto the inspection scene area in the scene 3D map.

[0026] The historical image data of various video data is converted to grayscale and a pixel threshold is set. Then, pixels with pixel values ​​greater than or equal to the pixel threshold in various historical image data are labeled, while pixels with pixel values ​​less than the pixel threshold are not labeled.

[0027] Based on the historical location information of the drones, the historical image data generated by drones that are located in the same spatial location and are performing the same drone inspection route are overlaid and mapped.

[0028] Set a threshold for the number of annotations. Based on the overlapping mapping results of historical image data, count the number of annotations for each pixel. Retain the annotations of pixels whose number of annotations is greater than or equal to the threshold, and discard the rest.

[0029] By connecting adjacent and labeled pixels, several feature regions can be divided in various historical image data.

[0030] Furthermore, the process of spatiotemporal alignment of historical image data includes:

[0031] Based on the location distribution of the feature regions, the corresponding feature regions are marked on the historical image data that has not been grayscaled.

[0032] Overlay and map various historical image data generated by the same drone in the same spatiotemporal sequence, use the scene content corresponding to the feature region as the positioning coordinate, spatially align historical image data generated in the same spatiotemporal sequence, and at the same time, use the method of spatially aligning historical image data collected by the same drone to spatially align historical image data generated by different drones at the same spatial location but executing the same drone inspection route.

[0033] Establish a three-dimensional coordinate system, map the different UAV inspection paths onto the three-dimensional coordinate system simultaneously, align the coordinates with the same parts, compare the different UAV inspection paths in the three-dimensional coordinate system in space, and retain the differences.

[0034] Based on the spatial alignment results of the UAV inspection paths that differ in the three-dimensional coordinate system, and the starting position of each UAV, the time nodes when each UAV collects the same target inspection area along the UAV inspection path but with different collection ranges are obtained and recorded as the synchronization difference time point group.

[0035] Then, the historical image data generated by drones performing different drone inspection paths under the same synchronization difference time point group are stitched together, and the different types of stitched historical image data are spatially aligned according to the distribution of feature regions.

[0036] Furthermore, the process of constructing a dynamic visualization scene model includes:

[0037] When the historical state data is generated and aligned under normal weather conditions and no abnormalities in the power system scenario, corresponding equipment contour models are established based on the feature regions in various historical video data in the historical state data. The three types of historical video data are then overlapped to obtain a multi-state equipment model.

[0038] Based on the target inspection area corresponding to the multi-state device model, the multi-state device model is mapped onto the scene 3D map to obtain a dynamic visualization scene model of the historical dynamic data for the corresponding number of days.

[0039] The dynamic visualization scene model is divided into several model voxels. Each model voxel is cube-shaped, and each face has an independent pixel value, which includes visible light pixels, infrared pixels and ultraviolet light pixels.

[0040] Then, the pixel values ​​of the same face of the model voxel from different days but corresponding to the same spatiotemporal order are distributed normally, and the center value of the normal distribution result is selected as the standard pixel value of each face of the corresponding model voxel under the corresponding spatiotemporal order.

[0041] Furthermore, the comparison process between the multi-state device model and the dynamic visualization scene model includes:

[0042] Starting from the first unit of time after each drone begins its inspection along the drone inspection path, after each unit of time, all real-time status data is divided into several real-time image data by frame, and defogging and shadow removal operations are performed on various real-time image data. A real-time local multi-state device model is constructed based on the real-time status data generated from the same frame.

[0043] The real-time local multi-state device model is divided into real-time model voxels of the same volume as the dynamic visualization scene model. The real-time model voxels are subtracted from the standard pixel values ​​of each face of the model voxels in spatiotemporal order, and three pixel difference thresholds are set according to the sensor type.

[0044] If the pixel difference of any pixel is greater than or equal to the pixel difference threshold, then an anomaly is determined to exist on one face of the corresponding real-time model voxel, and the type of pixel with the anomaly is marked; otherwise, the corresponding face is determined to be normal.

[0045] Furthermore, the process of locating abnormal locations includes:

[0046] If the same face of the same real-time model voxel is found to be abnormal based on the real-time status data of a complete unit of time, then the power equipment associated with the real-time model voxel is found to be abnormal, and the abnormal location is located based on the location of the real-time model voxel in the power equipment.

[0047] When an anomaly is detected in the power equipment, the system continuously monitors whether there is an anomaly in the real-time model voxels adjacent to the anomaly in the real-time model voxels, and determines the direction of anomaly propagation based on the monitoring results.

[0048] A machine vision-based intelligent inspection system for power systems includes a drone management and control module, an operation analysis module, and an equipment anomaly monitoring module.

[0049] The drone management module is used to acquire a GIS map of the power system scenario, set drone inspection paths based on the GIS map, perform flight simulations on the drone inspection paths, adjust the drone inspection paths according to the flight simulation results, and then the drones perform inspections in the power system scenario by adjusting the drone inspection paths and collect various real-time status data.

[0050] The operation analysis module is used to acquire historical state data of power system scenarios, divide the historical state data into historical image data, and mark multiple feature regions in the historical image data. Based on the distribution of feature regions, the historical image data is spatiotemporally aligned, and then a dynamic visualization scene model is constructed based on the historical image data.

[0051] The device anomaly monitoring module is used to establish a real-time local multi-state device model based on real-time status data, divide the real-time local multi-state device model and the dynamic visualization scene model into model voxels, and locate the anomaly location based on the model voxel comparison results.

[0052] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0053] 1. This invention systematically processes historical state data, including operations such as dividing feature regions and matching and fusing. Then, through grayscale processing, setting pixel thresholds and annotation frequency thresholds, feature regions are accurately divided. Based on the location distribution of feature regions, historical image data is matched, fused, and spatially aligned, so that different historical data can be effectively integrated, providing a reliable data foundation for building dynamic visualization scene models.

[0054] 2. This invention compares a real-time local multi-state device model with a dynamic visualization scene model to determine the abnormality of the real-time model voxels. Based on the type of abnormal pixels, it identifies the specific abnormality of the power equipment, such as appearance defects, overload, corona discharge, etc. At the same time, it accurately locates the abnormality based on the location of the real-time model voxels in the power equipment and determines the direction of abnormality propagation, providing strong support for timely handling of power equipment abnormalities. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0056] Figure 1 This is a flowchart of a machine vision-based intelligent inspection method for power systems according to the present invention.

[0057] Figure 2 This is a system framework diagram of an intelligent inspection system for power systems based on machine vision, as described in this invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Please see Figure 1 As shown, an intelligent inspection method for power systems based on machine vision includes the following steps:

[0060] Step S1: Obtain a GIS map of the power system scenario, set the drone inspection path based on the GIS map, conduct flight simulation of the drone inspection path, adjust the drone inspection path according to the flight simulation results, and then the drone will conduct inspection in the power system scenario by adjusting the drone inspection path and collect various real-time status data.

[0061] Step S2: Obtain historical state data of the power system scenario, divide the historical state data into historical image data, and mark multiple feature regions in the historical image data. Perform spatiotemporal alignment of the historical image data according to the distribution of feature regions, and then construct a dynamic visualization scenario model based on the historical image data.

[0062] Step S3: Establish a real-time local multi-state device model based on real-time status data, divide the real-time local multi-state device model and the dynamic visualization scene model into model voxels, and locate the abnormal location based on the model voxel comparison results.

[0063] Step S1 specifically includes the following process:

[0064] Step S101: Set the drone inspection route. The specific process includes:

[0065] Set up n drones in a power system scenario, with each drone starting at a different position in the power system scenario;

[0066] The drones are equipped with visible light imagers, infrared imagers, solar-blind ultraviolet detectors, and positioning devices, and each drone is assigned a number a1, a2, a3, ..., a n n is a natural number greater than 10;

[0067] Obtain a GIS map of the power system scenario, mark the patrol targets on the GIS map, such as towers, transformers, circuit breakers, insulator strings, etc., and build a 3D map of the scenario based on the GIS map;

[0068] Set a unit flight speed for the drone, and based on the flight distance of the unit flight speed in a unit time, divide the scene 3D map into several scene areas based on the flight distance. The unit time is generally 5 to 10 seconds, and the altitude value of each scene area is marked.

[0069] The scene area associated with each patrol target is taken as the target inspection area. According to the type of power equipment corresponding to each target inspection area, the no-fly radius is marked for each target inspection area. For example, the no-fly radius of high-voltage cables is 10m and the no-fly radius of transformers is 50m. This is used to reduce the electromagnetic interference generated by power equipment on the drone's sensors.

[0070] Set a fixed data collection range and fixed shooting angle for each drone, and then take the shortest path as the goal, the maximum inspection area coverage, the no-fly radius, and obstacle avoidance as prerequisites.

[0071] The shortest path is generated under the premise that the inspection path of the UAV is as short as possible.

[0072] The maximum inspection area coverage means that the drone's data collection range should include as many patrol targets in the power system scenario as possible within the drone's inspection path.

[0073] The obstacle avoidance means that the drone inspection path prioritizes avoiding the cruise target in the power system scenario and other scene objects (such as trees, houses, etc.) in order to avoid collisions during the drone's flight along the drone inspection path.

[0074] Using the starting position of each drone as the inspection start point and inspection end point, n drone inspection paths with different inspection start points and inspection end points but the same spatial location are generated in the scene 3D map;

[0075] Simultaneously conduct flight simulations of each UAV inspection path in the scene 3D map at a unit flight speed, and determine whether the inspection area of ​​each target in the scene 3D map is completely covered by the data collection range of the UAV based on the flight simulation results.

[0076] If it is determined that a target inspection area is not completely covered, then randomly select... The inspection path of the drone is shifted and adjusted. Since the drone has a fixed data acquisition range and a fixed shooting angle, the data acquisition range is also shifted and adjusted simultaneously after the inspection path is shifted and adjusted. This makes the flight simulation results of the adjusted drone inspection path cover the part of the data acquisition range that is not completely covered, but at the same time, there is some overlap with the original drone inspection path.

[0077] Repeat the process of judging and adjusting the drone inspection path until the inspection area of ​​each target in the scene 3D map is completely covered by the data collection range of the drone.

[0078] Step S102: Collect real-time scenario data of the power system scenario. The specific process includes:

[0079] The inspection paths of each drone are sent to the corresponding drones according to the inspection start point and inspection end point. Then, each drone conducts flight inspections in the power system scene at the same time according to the drone inspection path, and collects real-time status data of its power equipment through visible light imager, infrared imager and solar blind ultraviolet detector.

[0080] The real-time status data includes optical video data, infrared video data, and ultraviolet video data. At the end of each unit of time, each UAV collects the real-time video data and real-time location information collected by each sensor, and compresses and marks the real-time video data and real-time location information as real-time status data.

[0081] Step S2 specifically includes the following process:

[0082] Step S201: Divide the feature regions. The specific process includes:

[0083] Acquire historical status data of each power device in the power system scenario for several days under the condition that there are no abnormalities, and perform defogging and deshading operations on the historical status data of each day;

[0084] The historical status data collected by each drone is synchronized in time. The historical video data is divided into several historical image data by frame. Then, based on the historical location information and shooting angle of each drone, the historical image data corresponding to each drone is mapped onto the inspection scene area in the scene 3D map.

[0085] The historical image data of various video data is converted to grayscale and a pixel threshold is set. Then, pixels with pixel values ​​greater than or equal to the pixel threshold in various historical image data are labeled, while pixels with pixel values ​​less than the pixel threshold are not labeled.

[0086] Based on the historical location information of the drones, the historical image data generated by drones that are located in the same spatial location and are performing the same drone inspection route are overlaid and mapped.

[0087] Set a labeling frequency threshold. Based on the historical image data overlap mapping results, count the number of labels on each pixel. Keep the labels of pixels with a labeling frequency greater than or equal to the labeling frequency threshold, and discard the rest. The labeling frequency threshold is (n / 2, n).

[0088] Since the labeled pixels in the historical image data are all labeled by comparing with the same pixel threshold, and since the cruise targets in the power system scenario (such as towers, transformers, circuit breakers, insulator strings, etc.) will not change significantly within a certain period of time, the historical image data collected at the same spatial location at different times are overlapped with each other, and the occurrence frequency of labeled pixels is used to determine whether the image part associated with the pixel is a scene object that is not easy to change or the same scene object, thereby improving the accuracy of subsequent feature region segmentation.

[0089] By connecting adjacent and labeled pixels, several feature regions can be divided in various historical image data.

[0090] Step S202: Perform spatiotemporal alignment on historical image data. The specific process includes:

[0091] Based on the location distribution of the feature regions, the corresponding feature regions are marked on the historical image data that has not been grayscaled.

[0092] Since all kinds of historical video data are generated in the same spatiotemporal sequence and correspond to the same scene location, the historical video data collected by each sensor may have different data display formats, but the actual corresponding scene content is the same.

[0093] Overlay and map various historical image data generated by the same drone in the same spatiotemporal sequence, use the scene content corresponding to the feature region as the positioning coordinate, spatially align historical image data generated in the same spatiotemporal sequence, and at the same time, use the method of spatially aligning historical image data collected by the same drone to spatially align historical image data generated by different drones at the same spatial location but executing the same drone inspection route.

[0094] Establish a three-dimensional coordinate system and simultaneously map the different UAV inspection paths onto the three-dimensional coordinate system. Since each UAV inspection path only has some differences, there must be some common parts on each UAV inspection path. Then, align the coordinates with the common parts, compare the different UAV inspection paths in space in the three-dimensional coordinate system, and retain the differences.

[0095] Based on the spatial alignment results of the UAV inspection paths that differ in the three-dimensional coordinate system, and the starting position of each UAV, the time nodes when each UAV collects the same target inspection area along the UAV inspection path but with different collection ranges are obtained and recorded as the synchronization difference time point group.

[0096] Due to the inconsistency of drone inspection routes, the historical status data collected by some drones under the same conditions may differ. However, since the target power equipment collected by drones in the same spatial location is the same, and there is some overlap before and after the drone inspection path is adjusted, the collection results may be partially the same.

[0097] Then, the historical image data generated by drones performing different drone inspection paths under the same synchronization difference time point group are stitched together, and the different types of stitched historical image data are spatially aligned according to the distribution of feature regions.

[0098] Step S203: Construct a dynamic visualization scene model. The specific process includes:

[0099] When the historical state data is generated and aligned under normal weather conditions and no abnormalities in the power system scenario, corresponding equipment contour models are established based on the feature regions in various historical video data in the historical state data. The three types of historical video data are then overlapped to obtain a multi-state equipment model.

[0100] Based on the target inspection area corresponding to the multi-state device model, the multi-state device model is mapped onto the scene 3D map to obtain a dynamic visualization scene model of the historical dynamic data for the corresponding number of days.

[0101] The dynamic visualization scene model is divided into several model voxels. Each model voxel is cube-shaped, and each face has an independent pixel value, which includes visible light pixels, infrared pixels and ultraviolet light pixels.

[0102] Then, the pixel values ​​of the same face of the model voxel from different days but corresponding to the same spatiotemporal order are distributed normally, and the center value of the normal distribution result is selected as the standard pixel value of each face of the corresponding model voxel under the corresponding spatiotemporal order.

[0103] Step S3 specifically includes the following process:

[0104] Step S301, anomaly detection, specifically includes:

[0105] Starting from the first unit of time after each drone begins its inspection along the drone inspection path, after each unit of time, all real-time status data is divided into several real-time image data by frame, and defogging and shadow removal operations are performed on various real-time image data. A real-time local multi-state device model is constructed based on the real-time status data generated from the same frame.

[0106] The real-time local multi-state device model is divided into real-time model voxels of the same volume as the dynamic visualization scene model. The real-time model voxels are subtracted from the standard pixel values ​​of each face of the model voxels in spatiotemporal order, and three pixel difference thresholds are set according to the sensor type.

[0107] If the pixel difference of any pixel is greater than or equal to the pixel difference threshold, then it is determined that there is an anomaly on one face of the corresponding real-time model voxel, and the type of pixel with the anomaly is marked.

[0108] If the pixel difference of all types of pixels is less than the pixel difference threshold, then the corresponding surface is considered normal.

[0109] Step S302: Locate the abnormal location. The specific process includes:

[0110] If the same face of the same real-time model voxel is found to be abnormal based on the real-time status data of a complete unit of time, then the power equipment associated with the real-time model voxel is found to be abnormal, and the abnormal location is located based on the location of the real-time model voxel in the power equipment.

[0111] If the visible light pixels are abnormal, it is determined that the corresponding power equipment has external defects. If the infrared pixels are abnormal, it is determined that the corresponding power equipment is overloaded or has stopped operating. If the ultraviolet pixels are abnormal, it is determined that the corresponding power equipment has corona discharge.

[0112] When an anomaly is detected in the power equipment, the system continuously monitors whether there is an anomaly in the real-time model voxels adjacent to the anomaly in the real-time model voxels, and determines the direction of anomaly propagation based on the monitoring results.

[0113] Please see Figure 2 As shown, an intelligent inspection system for power systems based on machine vision includes a drone management and control module, an operation analysis module, and an equipment anomaly monitoring module.

[0114] The drone management module is used to acquire a GIS map of the power system scenario, set drone inspection paths based on the GIS map, perform flight simulations on the drone inspection paths, adjust the drone inspection paths according to the flight simulation results, and then the drones perform inspections in the power system scenario by adjusting the drone inspection paths and collect various real-time status data.

[0115] The operation analysis module is used to acquire historical state data of power system scenarios, divide the historical state data into historical image data, and mark multiple feature regions in the historical image data. Based on the distribution of feature regions, the historical image data is spatiotemporally aligned, and then a dynamic visualization scene model is constructed based on the historical image data.

[0116] The device anomaly monitoring module is used to establish a real-time local multi-state device model based on real-time status data, divide the real-time local multi-state device model and the dynamic visualization scene model into model voxels, and locate the anomaly location based on the model voxel comparison results.

[0117] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent inspection of power system based on machine vision, characterized in that, The method comprises the following steps: S1, acquiring a GIS map of the power system scene, setting a UAV inspection path based on the GIS map, simulating flight of the UAV inspection path, adjusting the UAV inspection path according to the flight simulation result, and then having the UAV inspect the power system scene through the adjusted UAV inspection path and collect various real-time state data; The real-time state data comprises optical video data, infrared video data and ultraviolet video data, and the real-time state data is generated by integrating the real-time video data collected by the sensors and the real-time position information of each UAV every time a unit time ends; S2, acquiring historical state data of the power system scene, dividing the historical state data into historical image data, marking a plurality of feature regions in the historical image data, temporally and spatially aligning the historical image data according to the feature region distribution, and then constructing a dynamic visual scene model according to the historical image data; The process of marking a plurality of feature regions in the historical image data comprises: Acquiring historical state data of the power system scene in which each power equipment is normal for several days, dividing each historical video in the historical state data into several parts of historical image data according to frames, and mapping the historical image data on a scene three-dimensional map according to the historical position information and the shooting angle of each UAV; Performing grayscale processing on the historical image data, setting a pixel threshold, marking pixels with a pixel value greater than or equal to the pixel threshold in the historical image data, and not marking pixels with a pixel value less than the pixel threshold; According to the historical position information of the UAV, the historical image data generated by the UAVs located at the same spatial position and executing the same UAV inspection route are overlapped and mapped; Setting a marking frequency threshold, counting the marking frequency of each pixel according to the overlapped and mapped results of the historical image data, retaining the marking of pixels with a marking frequency greater than or equal to the marking frequency threshold, and otherwise eliminating, connecting adjacent pixels with markings, and dividing a plurality of feature regions in each historical image data; The construction process of the dynamic visual scene model comprises: Establishing a device contour model according to the feature regions in various historical video data in the historical state data, and overlapping the device contour models corresponding to the three kinds of historical video data to obtain a multi-state device model; Mapping the multi-state device model in the scene three-dimensional map to obtain a dynamic visual scene model according to the target inspection region corresponding to the multi-state device model, dividing the dynamic visual scene model into a plurality of model voxels, and the model voxels are in the shape of a cube, each face has an independent pixel value, and simultaneously contains a visible light pixel, an infrared pixel and an ultraviolet light pixel; Normal distribution is performed on the pixel values of the same faces of the model voxels from different days but corresponding to a time-space sequence, and the central value of the normal distribution result is selected as the standard pixel value of each face of the model voxel in the time-space sequence; S3, establishing a real-time local multi-state device model according to the real-time state data, dividing the real-time local multi-state device model and the dynamic visual scene model into model voxels, and positioning an abnormal position according to the comparison result of the model voxels. 2.The method of claim 1, wherein, The setting process of the UAV inspection route comprises: n unmanned aerial vehicles are arranged in the power system scene, and starting positions of the unmanned aerial vehicles in the power system scene are different from each other; The respective unmanned aerial vehicles are provided with numbers a1, a2, a3,..., a n n is a natural number greater than 10; a GIS map of the power system scene is acquired, and a cruising target is marked on the GIS map, and a three-dimensional map of the scene is established based on the GIS map; a unit flight speed is set for the unmanned aerial vehicles, a flight distance of the unit flight speed in a unit time is determined, and then the three-dimensional map of the scene is divided into a plurality of scene regions according to the flight distance, and a height value of each scene region is marked; a scene region associated with each cruising target is taken as a target inspection region, a flight radius is marked for each target inspection region according to a type of power equipment corresponding to the target inspection region; a fixed data acquisition range and a fixed shooting angle are set for each unmanned aerial vehicle, and then, under the precondition of a shortest path, a maximum inspection region coverage, a flight radius and obstacle avoidance, starting positions of the unmanned aerial vehicles are taken as a starting point and an end point of inspection, and n unmanned aerial vehicle inspection paths are generated on the three-dimensional map of the scene, the starting points and the end points of the unmanned aerial vehicle inspection paths are different, and spatial positions passed through by the unmanned aerial vehicle inspection paths are the same. 3.The method of claim 2, wherein, a flight simulation process of the unmanned aerial vehicle inspection path includes: flight simulation is performed on each unmanned aerial vehicle inspection path in the three-dimensional map of the scene at the unit flight speed, and whether each target inspection region in the three-dimensional map of the scene is completely covered by the data acquisition range of the unmanned aerial vehicle is determined according to a flight simulation result; If it is judged that the target inspection area is not completely covered, randomly adjust The unmanned aerial vehicle inspection path is adjusted so that the flight simulation result of the adjusted unmanned aerial vehicle inspection path covers the part of the data collection range that is not completely covered, but at the same time there is partial overlap with the original unmanned aerial vehicle inspection path. 4.The method of claim 3, wherein, a real-time scene data acquisition process of the power system scene includes: each unmanned aerial vehicle inspection path is sent to the unmanned aerial vehicle according to the starting point and the end point of inspection, and then each unmanned aerial vehicle flies and inspects in the power system scene according to the unmanned aerial vehicle inspection path, and real-time state data of power equipment is acquired; the real-time state data includes optical video data, infrared video data and ultraviolet video data, and each time a unit time ends, real-time video data and real-time position information acquired by each unmanned aerial vehicle are integrated to generate the real-time state data.

5. The method of claim 4, wherein, a spatio-temporal alignment process of historical image data includes: according to a feature region position distribution, corresponding feature regions are marked on historical image data that has not been subjected to grayscale processing; various historical image data generated by the same unmanned aerial vehicle in the same spatio-temporal sequence are overlapped and mapped, a scene content corresponding to a feature region is taken as a positioning coordinate, historical image data generated in the same spatio-temporal sequence is spatially aligned, and a method of spatially aligning historical image data acquired by the same unmanned aerial vehicle is used to spatially align historical image data generated by different unmanned aerial vehicles in the same spatial position but executing the same unmanned aerial vehicle inspection route; the different unmanned aerial vehicle inspection paths are simultaneously mapped on a three-dimensional coordinate system, and then the different unmanned aerial vehicle inspection paths are spatially compared in the three-dimensional coordinate system by taking the same part as an alignment coordinate, and a difference part is retained; according to a spatial alignment result of the different unmanned aerial vehicle inspection paths in the three-dimensional coordinate system and starting positions of the unmanned aerial vehicles, time nodes at which the same target inspection region is acquired by each unmanned aerial vehicle along the unmanned aerial vehicle inspection path but the acquisition ranges are different are acquired, and the time nodes are recorded as a synchronous difference time point group. Further, the unmanned aerial vehicles performing different unmanned aerial vehicle inspection paths are spliced with historical image data generated at the same synchronization difference point group, and different types of spliced historical image data are spatially aligned according to feature region distribution.

6. The method of claim 5, wherein the method further comprises: The comparison process of the multi-state device model and the dynamic visual scene model includes: After each unmanned aerial vehicle starts inspection along the unmanned aerial vehicle inspection path, the real-time state data is divided into real-time image data in each frame after each unit time, the real-time image data is defogged and de-shadowed, and the real-time local multi-state device model is constructed according to the real-time state data generated in the same frame; The real-time local multi-state device model is divided into real-time model voxels with the same volume as the dynamic visual scene model, the real-time model voxels are subtracted from the standard pixel value of each face of the model voxels in time and space order, and three pixel difference thresholds are set according to the sensor type; If the pixel difference of any pixel is greater than or equal to the pixel difference threshold, it is determined that there is an abnormality in one face of the corresponding real-time model voxel, and the pixel type of the abnormal pixel is marked, otherwise it is determined that the corresponding face is normal.

7. The method of claim 6, wherein the method further comprises: The process of locating the abnormal position includes: If the same face of the same real-time model voxel is determined to be abnormal according to the real-time state data of a complete unit time, it is determined that the power equipment associated with the real-time model voxel is abnormal, and the abnormal position is located according to the position of the real-time model voxel in the power equipment; When it is determined that the power equipment is abnormal, it is continuously monitored whether the real-time model voxels adjacent to the abnormal real-time model voxel are abnormal, and the abnormal spread direction is determined according to the monitoring result.

8. An intelligent inspection system of power system based on machine vision, used to implement the intelligent inspection method of power system based on machine vision in any one of claims 1-7, characterized in that, The unmanned aerial vehicle control module, the operation analysis module, and the equipment abnormality monitoring module are included. The unmanned aerial vehicle control module is used to obtain the GIS map of the power system scene, set the unmanned aerial vehicle inspection path based on the GIS map, simulate the flight of the unmanned aerial vehicle inspection path, adjust the unmanned aerial vehicle inspection path according to the flight simulation result, and then the unmanned aerial vehicle inspects the power system scene by adjusting the unmanned aerial vehicle inspection path and collects multiple real-time state data. The operation analysis module is used to obtain the historical state data of the power system scene, divide the historical state data into historical image data, mark multiple feature regions in the historical image data, spatially align the historical image data according to the feature region distribution, and then construct a dynamic visual scene model according to the historical image data. The equipment abnormality monitoring module is used to establish a real-time local multi-state device model according to the real-time state data, divide the real-time local multi-state device model and the dynamic visual scene model into model voxels, and locate the abnormal position according to the comparison result of the model voxels.

Images