A power transmission line video flashover fault detection method and system based on dynamic baseline multi-dimensional feature scoring

Abstract

The application discloses a power transmission line video flashover fault detection method and system based on dynamic baseline multi-dimensional feature scoring, and relates to the technical field of intelligent diagnosis of power transmission line faults. The method acquires a power transmission line monitoring video, extracts overall brightness, white proportion, highlight proportion, texture complexity and regional features; establishes a dynamic baseline according to historical video frames and determines a dynamic threshold in combination with a standard deviation, scores abnormal features to determine flashover candidate frames; performs duration filtering and recovery mode checking on the flashover candidate frames to exclude persistent changes; when no abnormal frame is detected at the frame level, performs gradual offset supplementary detection according to white proportion baseline offset, and outputs a flashover fault detection result. The application can improve the detection adaptability of local, instantaneous and weak flashover events under dynamic lighting scenes, and reduce false positives and missed detections.

Description

Technical Field

[0001] This invention belongs to the field of intelligent fault diagnosis technology for power transmission lines, and particularly relates to a method and system for detecting flash explosion faults in power transmission lines based on dynamic baseline multidimensional feature scoring. Background Technology

[0002] Transmission lines are a crucial component of power systems. When transmission lines experience faults such as short circuits, lightning strikes, insulator flashovers, or tree-line discharges, abnormal optical phenomena such as electric arcs and flash explosions may occur. With the gradual deployment of online monitoring equipment for transmission lines, monitoring videos can record changes in the image during transmission line operation, providing a video data foundation for detecting abnormal optical events when faults occur.

[0003] Currently, abnormal optical events such as electric arcs and flashes in power transmission line monitoring videos are typically detected using methods such as inter-frame differential and fixed thresholds. However, power transmission line monitoring scenes are usually located in outdoor environments, and the images are easily affected by changes in ambient lighting such as day and night cycles, cloud cover, and vehicle headlights. When using fixed thresholds for detection, it is difficult to adjust the detection conditions according to different video scenes and image fluctuations at different times, which can easily lead to false positives or false negatives of flashes.

[0004] Meanwhile, electric arcs and flashes occurring on power transmission lines are typically characterized by short duration and momentary brightening of localized areas. Some flash events last only a few video frames. When using global features such as overall brightness and overall pixel ratio for detection, abnormal changes in localized areas are easily weakened by the statistical results of the entire video frame, making it difficult to effectively identify localized flash events.

[0005] In addition, scene transitions, gradual changes in lighting, and automatic camera parameter adjustments can also cause changes in brightness or color in video footage. These changes are usually continuous or gradual, differing from the instantaneous changes caused by flash explosions. However, judging solely based on changes in a single frame or between adjacent video frames is insufficient to effectively distinguish flash explosion events from these normal scene changes, easily leading to false alarms.

[0006] On the other hand, some weak flash explosion events do not exhibit obvious single-frame jumps, but rather manifest as a baseline shift in the proportion of white in the video frame at a certain moment, which continues to change in subsequent video frames. Traditional inter-frame difference detection methods mainly rely on instantaneous changes between adjacent video frames for judgment, which carries the risk of missing detection for such weak flash explosion events without significant single-frame jumps.

[0007] Therefore, how to detect short-term flash explosions, local flash explosions, and weak flash explosions without significant single-frame jumps in the monitoring scenario of power transmission lines, and reduce false alarms caused by dynamic lighting and continuous image changes, is a technical problem that needs to be solved in the video flash explosion fault detection of power transmission lines. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention proposes a method and system for detecting flash explosion faults in transmission lines based on dynamic baseline multidimensional feature scoring.

[0009] The technical solution of the present invention is as follows:

[0010] A method for detecting flashover faults in transmission lines based on dynamic baseline multidimensional feature scoring includes:

[0011] Acquire monitoring videos of power transmission lines and extract multidimensional visual features from each video frame. The multidimensional visual features include overall brightness, white ratio, highlight ratio, texture complexity, and regional features.

[0012] Based on the multidimensional visual features of historical video frames, corresponding dynamic baselines are established and their standard deviations are calculated. Dynamic thresholds are determined based on preset absolute thresholds and the standard deviations. Abnormal features are determined based on the comparison between the change of the multidimensional visual features of the current frame relative to the corresponding dynamic baseline and the corresponding dynamic threshold. The abnormal features are scored, and flash explosion candidate frames are determined based on the comprehensive scores.

[0013] Temporally consecutive flash burst candidate frames are merged into events. Duration filtering is performed based on the duration of each event. The flash burst candidate frames retained after duration filtering are then subjected to recovery mode checks to exclude persistent changes, thus obtaining frame-level detection results.

[0014] When no abnormal frame is detected in the frame-level detection results, a gradual offset supplementary detection is performed on the white proportion sequence of the video frame, and the flash explosion event and the corresponding abnormal frame are determined based on the offset of the white proportion baseline.

[0015] Based on the frame-level detection results or the results of the gradient offset supplementary detection, the flashover fault detection results of the transmission line video are output.

[0016] Furthermore, the overall brightness is obtained by converting the video frame into a grayscale image and calculating the grayscale mean; the white proportion is obtained by converting the video frame to the HSV color space and calculating the proportion of pixels that simultaneously meet preset low saturation and preset high brightness conditions to the total pixels of the video frame; the high brightness proportion is obtained by thresholding the grayscale image of the video frame and calculating the proportion of pixels exceeding the preset high brightness threshold to the total pixels of the video frame; the texture complexity is obtained by calculating the Laplacian variance of the grayscale image of the video frame; the multidimensional visual features also include the red proportion, which is obtained by detecting red pixels in the two hue ranges of 0°~10° and 160°~180° in the HSV color space that meet preset saturation and preset brightness conditions, and calculating the proportion of red pixels to the total pixels of the video frame.

[0017] Furthermore, the extraction of the regional features includes dividing the video frame into a 4×4 grid, calculating the local brightness, local white ratio, and local highlight ratio in each grid, and determining the maximum white ratio and maximum highlight ratio of the region based on the calculation results corresponding to each grid.

[0018] Furthermore, the step of establishing corresponding dynamic baselines and calculating standard deviations based on the multidimensional visual features of historical video frames includes establishing and maintaining a feature history queue with a sliding window of a preset length, calculating the median and standard deviation of each dimension of visual features in the feature history queue, and using the median as the dynamic baseline of the corresponding visual feature; for each dimension of visual features of the current frame, taking the larger value between the corresponding preset absolute threshold and twice the corresponding standard deviation as the dynamic threshold.

[0019] Furthermore, when scoring each abnormal feature, 2 points are awarded when the white proportion suddenly increases beyond the corresponding dynamic threshold, 2 points are awarded when the highlight proportion suddenly increases beyond the corresponding dynamic threshold, 1 point is awarded when the overall brightness suddenly increases beyond the corresponding dynamic threshold, 1 point is awarded when the texture complexity decreases beyond the preset texture decrease threshold, 2 points are awarded when the growth rate of the maximum white proportion in a region relative to the corresponding dynamic baseline exceeds 50%, and 2 points are awarded when the growth rate of the maximum highlight proportion in a region relative to the corresponding dynamic baseline exceeds 50%. The preset score threshold for the comprehensive score is 3 points.

[0020] Furthermore, supplementary detection is performed on the current frame based on the red proportion. When the growth rate of the red proportion of the current frame relative to the dynamic baseline corresponding to the red proportion exceeds 100%, and the overall brightness of the current frame is higher than the dynamic baseline corresponding to the overall brightness or the absolute value of the overall brightness exceeds a preset brightness threshold, the current frame is determined as a flash explosion candidate frame.

[0021] Furthermore, the duration filtering includes recording the duration of each event through an event state machine. When the duration of an event exceeds 20 frames, only the flash burst candidate frame detection results corresponding to the first 5 frames of the event are retained. When the number of flash burst candidate frames in an event is less than 2 frames and there is a lack of color change evidence, all flash burst candidate frame detection results in the event are cleared. The color change evidence includes a sudden increase in the white proportion of the flash burst candidate frame accompanied by overexposure evidence, or a sudden increase in the maximum highlight proportion of the flash burst candidate frame area.

[0022] Furthermore, the recovery mode check includes grouping the flash explosion candidate frames retained after duration filtering into event clusters according to the condition that the frame interval between adjacent flash explosion candidate frames is no more than 2 frames, determining the white percentage event peak value of each event cluster, obtaining the white percentage value within the recovery window after each event cluster, judging whether the corresponding event cluster has recovery characteristics after instantaneous flash explosion based on the decrease magnitude between the white percentage event peak value and the white percentage value within the recovery window, and excluding event clusters whose white percentage value maintains a continuous shift and does not exhibit the recovery characteristics.

[0023] Furthermore, the gradient offset supplementation detection includes pre-checking the change in white proportion of adjacent video frames, performing baseline offset detection based on the median difference in white proportion within the window before and after the video, scanning forward video frames when baseline offset is detected, and determining the first video frame that meets the preset offset start point determination condition as the offset start point, verifying the initial change based on the inter-frame change in white proportion within the window before and after the offset start point, verifying the offset persistence based on the situation where the white proportion after the offset start point is continuously higher than the baseline before the offset, and performing color evidence verification. After all verifications are passed, the flash explosion start frame is determined based on the offset start point, and the video frames corresponding to the flash explosion start frame and a preset number of frames before and after it are marked as abnormal frames.

[0024] A transmission line video flash explosion fault detection system based on dynamic baseline multidimensional feature scoring includes a video input module, a feature extraction module, a dynamic baseline module, a detection and judgment module, a gradual offset module, and a result output module.

[0025] The video input module is used to acquire monitoring video of the power transmission line;

[0026] The feature extraction module is used to extract multi-dimensional visual features of video frames frame by frame. The multi-dimensional visual features include overall brightness, white ratio, highlight ratio, texture complexity, and region features.

[0027] The dynamic baseline module is used to establish corresponding dynamic baselines and calculate standard deviations based on the multidimensional visual features of historical video frames.

[0028] The detection and judgment module is used to determine the dynamic threshold based on the preset absolute threshold and the standard deviation, determine the abnormal features based on the comparison result of the change of the multidimensional visual features of the current frame relative to the corresponding dynamic baseline and the corresponding dynamic threshold, score the abnormal features and determine flash explosion candidate frames based on the comprehensive score, merge the temporally continuous flash explosion candidate frames into events, perform duration filtering based on the duration of each event, and perform recovery mode check on the flash explosion candidate frames retained after duration filtering to exclude continuous changes and obtain frame-level detection results.

[0029] The gradient offset module is used to perform gradient offset supplementary detection on the white proportion sequence of the video frame when the frame-level detection result does not detect abnormal frames, and to determine the flash explosion event and the corresponding abnormal frame based on the offset of the white proportion baseline.

[0030] The result output module is used to output the flashover fault detection result of the transmission line video based on the frame-level detection result or the detection result of the gradient offset module.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] This invention extracts the overall brightness, white ratio, highlight ratio, texture complexity, and regional features of video frames, and performs a comprehensive score based on the anomalies of each visual feature to determine candidate frames for flashover. By characterizing the abnormal state of video frames from multiple aspects such as brightness, color, overexposure, texture, and local regional changes, it can reduce the problem of being easily affected by local noise or fluctuations in a single feature when judging flashover based on only a single visual feature, thereby improving the ability to identify flashover events in transmission line videos.

[0033] This invention establishes a corresponding dynamic baseline based on the multidimensional visual features of historical video frames and determines a dynamic threshold by combining the standard deviation of each visual feature. This allows the anomaly judgment conditions of the current frame to be adjusted according to the fluctuation of historical video images, solving the problem that fixed detection thresholds are difficult to adapt to different lighting environments and scene fluctuations. Thus, it retains the ability to detect small abnormal changes when the image is relatively stable, and reduces the probability of normal fluctuations being judged as flash explosion events when the image fluctuates greatly, thereby improving the adaptability of flash explosion detection to dynamic monitoring scenarios.

[0034] This invention divides video frames into multiple grid regions, extracting local brightness, local white percentage, and local highlight percentage within each grid. The maximum white percentage and maximum highlight percentage of each region are then used for flash detection. This avoids the weakening of instantaneous flashes in local areas during global statistical analysis, solves the problem of insensitivity to local flash events when only overall image features are used, and improves the detection capability of local flash events.

[0035] This invention merges temporally consecutive flash explosion candidate frames into events, performs duration filtering based on the event's duration, and then checks the recovery mode of the remaining flash explosion candidate frames after duration filtering. It determines whether the flash explosion is instantaneous or a continuous change based on whether relevant features are restored after the flash explosion event. This allows for further differentiation of images with continuous changes, such as scene transitions, gradual changes in lighting, and automatic camera parameter adjustments, reducing the probability of such normal scene changes being misjudged as flash explosion malfunctions.

[0036] When frame-level detection fails to detect abnormal frames, this invention further performs a gradual offset supplementary detection on the white proportion sequence of video frames. The offset starting point is determined based on the offset of the white proportion baseline before and after the frame, and the continuous changes after the offset are verified. This enables supplementary detection of weak flash explosion events where there are no significant single-frame jumps but the white proportion baseline shifts. It compensates for the problem of easily missing such flash explosion events when relying solely on instantaneous changes between frames, and improves the coverage of flash explosion events with different manifestations. Attached Figure Description

[0037] Figure 1 A schematic diagram of the video flash explosion fault detection method for transmission lines;

[0038] Figure 2 This is a schematic diagram of the recovery curve for a transient flash explosion event;

[0039] Figure 3 This is a schematic diagram of an unrecoverable curve representing a continuous change. Detailed Implementation

[0040] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0041] Example 1:

[0042] This invention provides a method for detecting flashover faults in transmission lines based on dynamic baseline multidimensional feature scoring, such as... Figure 1 As shown, it includes:

[0043] Acquire monitoring videos of power transmission lines and extract multi-dimensional visual features from each video frame. These multi-dimensional visual features include overall brightness, white ratio, highlight ratio, texture complexity, and regional features.

[0044] Based on the multidimensional visual features of historical video frames, establish corresponding dynamic baselines and calculate the standard deviation. Determine the dynamic threshold based on the preset absolute threshold and standard deviation. Determine abnormal features based on the comparison between the change of the multidimensional visual features of the current frame relative to the corresponding dynamic baseline and the corresponding dynamic threshold. Score the abnormal features and determine flash explosion candidate frames based on the comprehensive score.

[0045] Temporally consecutive flash burst candidate frames are merged into events. Duration filtering is performed based on the duration of each event. The flash burst candidate frames retained after duration filtering are then subjected to recovery mode checks to exclude persistent changes, thus obtaining frame-level detection results.

[0046] When no abnormal frames are detected in the frame-level detection results, a gradual offset supplementary detection is performed on the white proportion sequence of the video frames. Based on the offset of the white proportion baseline, the flash explosion event that did not produce a significant single-frame jump and the corresponding abnormal frame are determined.

[0047] Based on the frame-level detection results or the results of the gradual offset supplementary detection, the flashover fault detection results of the transmission line video are output.

[0048] Specifically, after obtaining abnormal frames, they can be divided into different event groups according to their temporal continuity. For each event, the first frame and the two frames before and after it are saved, and a structured detection report is generated. For example, the structured detection report can record the detection results, the range of abnormal frames, and the information on the saved abnormal frame images in JSON format.

[0049] Example 2:

[0050] This embodiment, based on Embodiment 1, further designs the following: In this example, the overall brightness is obtained by converting the video frame to a grayscale image and calculating the grayscale mean; the white proportion is obtained by converting the video frame to the HSV color space and calculating the proportion of pixels that simultaneously meet preset low saturation and preset high brightness conditions to the total pixels of the video frame; the high brightness proportion is obtained by thresholding the grayscale image of the video frame and calculating the proportion of pixels exceeding a preset high brightness threshold to the total pixels of the video frame; the texture complexity is obtained by calculating the Laplacian variance of the grayscale image of the video frame; the multidimensional visual features also include the red proportion, which is obtained by detecting red pixels in the two hue ranges of 0°~10° and 160°~180° in the HSV color space that meet preset saturation and preset brightness conditions, and calculating the proportion of red pixels to the total pixels of the video frame.

[0051] Specifically, the low saturation condition for white proportion is set to a saturation S≤30 in the HSV color space, and the high brightness condition is set to a brightness V≥200; the high brightness threshold for high brightness proportion is set to 230. For red proportion, pixels with hues in the ranges of 0°~10° and 160°~180° are detected respectively, and the lower limit of saturation for red pixels is set to 50 and the lower limit of brightness is set to 100. Pixels that meet the above saturation and brightness conditions are included in the red proportion calculation. Texture complexity is represented by the Laplacian variance of the grayscale image. When flash exposure causes local or overall overexposure, the texture information in the corresponding area decreases, and the Laplacian variance decreases accordingly.

[0052] Example 3:

[0053] This embodiment is further designed based on the first embodiment. In this example, the extraction of regional features includes dividing the video frame into a 4×4 grid, calculating the local brightness, local white ratio, and local highlight ratio in each grid, and determining the maximum white ratio and the maximum highlight ratio of the region based on the calculation results corresponding to each grid.

[0054] Specifically, after calculating the local brightness, local white ratio, and local highlight ratio of each grid, the grid position corresponding to the maximum value of each regional feature can be recorded simultaneously to characterize the spatial position of abnormal brightness or overexposed areas in the video frame, providing a positional reference for subsequent abnormal frame viewing and flash explosion area location.

[0055] Example 4:

[0056] This embodiment is further designed based on Embodiment 1. In this example, dynamic baselines are established and standard deviations are calculated according to the multidimensional visual features of historical video frames. This includes establishing and maintaining a feature history queue with a sliding window of a preset length, calculating the median and standard deviation of each dimension of visual features in the feature history queue, and using the median as the dynamic baseline of the corresponding visual feature. For each dimension of visual features of the current frame, the larger value between the corresponding preset absolute threshold and twice the corresponding standard deviation is taken as the dynamic threshold.

[0057] For example, the preset length of the sliding window can be set to 30 frames, that is, a feature history queue is established based on the visual features corresponding to 30 consecutive video frames. As video frames are continuously input, the feature history queue is updated by sliding, and the median and standard deviation of each dimension of visual features are recalculated. Among them, using the median as a dynamic baseline can reduce the impact of a small number of outlier feature values ​​on the baseline calculation results.

[0058] Example 5:

[0059] This embodiment, based on Embodiment 3, is further designed in that, when scoring each abnormal feature, a sudden increase in the proportion of white exceeding the corresponding dynamic threshold is scored as follows: 2 points are scored when the proportion of highlights suddenly exceeds the corresponding dynamic threshold; 1 point is scored when the overall brightness suddenly exceeds the corresponding dynamic threshold; 1 point is scored when the texture complexity decreases beyond the preset texture decrease threshold; 2 points are scored when the growth rate of the largest white proportion in a region relative to the corresponding dynamic baseline exceeds 50%; 2 points are scored when the growth rate of the largest highlight proportion in a region relative to the corresponding dynamic baseline exceeds 50%; and the preset score threshold for the comprehensive score is 3 points.

[0060] Specifically, a sudden increase in white proportion is used to characterize the white or bright white areas generated instantaneously by the flash, a sudden increase in highlight proportion is used to characterize the expansion of overexposed areas in the video frame, a sudden increase in overall brightness is used to characterize the instantaneous change in the overall lighting state of the image, and a decrease in texture complexity is used to characterize the reduction of texture information caused by overexposure; the maximum white proportion and the maximum highlight proportion of the region are used to characterize the white overexposure or bursting phenomenon within a local grid area. Thus, a comprehensive judgment is made on flash candidate frames through a combination of scores of different visual features.

[0061] Example 6:

[0062] This embodiment is further designed based on embodiment two. In this example, the current frame is also supplemented by detection based on the proportion of red. When the growth rate of the proportion of red in the current frame relative to the dynamic baseline corresponding to the proportion of red exceeds 100%, and the overall brightness of the current frame is higher than the dynamic baseline corresponding to the overall brightness or the absolute value of the overall brightness exceeds the preset brightness threshold, the current frame is determined as a flash explosion candidate frame.

[0063] Specifically, the red proportion supplementary detection serves as a decision path parallel to the multidimensional visual feature comprehensive scoring. When the comprehensive score of the current frame does not reach the preset score threshold, but the growth rate of the red proportion relative to the corresponding dynamic baseline exceeds 100% and meets the overall brightness increase condition, the current frame is still identified as a flash explosion candidate frame to supplement the detection of flash explosion events with obvious red features but weak changes in other visual features.

[0064] Example 7:

[0065] This embodiment, based on Embodiment 1, is further designed in that the duration filtering in this example includes recording the duration of each event through an event state machine. When the duration of an event exceeds 20 frames, only the flash burst candidate frame detection results corresponding to the first 5 frames of the event are retained. When the number of flash burst candidate frames in an event is less than 2 frames and there is a lack of color change evidence, all flash burst candidate frame detection results in the event are cleared. Color change evidence includes a sudden increase in the white proportion of the flash burst candidate frame accompanied by overexposure evidence, or a sudden increase in the maximum highlight proportion of the flash burst candidate frame area.

[0066] Specifically, color change evidence is used to characterize a significant abrupt increase in the white proportion and / or highlight proportion of a flash burst candidate frame relative to the baseline. This includes: a sudden increase in the white proportion of the flash burst candidate frame accompanied by overexposure evidence, or a sudden increase in the maximum highlight proportion of a region within the flash burst candidate frame; where overexposure evidence includes an increase in highlight proportion or a sudden increase in the maximum highlight proportion of a region. For example, a sudden increase in white proportion can be determined by its growth rate relative to the corresponding dynamic baseline exceeding a corresponding dynamic threshold, or by its absolute increase exceeding a preset white increase threshold (e.g., 0.015); an increase in highlight proportion can be determined by its absolute increase exceeding a preset highlight increase threshold (e.g., 0.01). The event state machine counts the number of frames of event duration based on temporally consecutive flash burst candidate frames. Events lasting 1 to 20 frames are treated as instantaneous flash burst events for subsequent detection, while events lasting longer than 20 frames are filtered as persistent changes.

[0067] Example 8:

[0068] This embodiment, based on Embodiment 1, further designs the recovery mode check as follows: the flash explosion candidate frames retained after duration filtering are grouped into event clusters according to the condition that the frame interval between adjacent flash explosion candidate frames is no more than 2 frames. The white percentage event peak value of each event cluster is determined, and the white percentage value within the recovery window after each event cluster is obtained. Based on the decrease in the white percentage event peak value and the white percentage value within the recovery window, it is determined whether the corresponding event cluster has recovery characteristics after an instantaneous flash explosion. Event clusters in which the white percentage value maintains a continuous shift and does not show recovery characteristics are excluded.

[0069] Specifically, the recovery window is the video frames corresponding to at most 8 frames after the event cluster ends; when there are fewer than 8 video frames remaining after the event cluster ends, the recovery window is all remaining video frames. During the determination, the difference between the peak white percentage of the event cluster and the minimum white percentage within the recovery window is calculated. When the difference is greater than a preset recovery gap threshold (e.g., 0.005), the corresponding event cluster is considered to have shown a significant decline after the flash burst, exhibiting recovery characteristics after an instantaneous flash burst, and is thus retained.

[0070] like Figure 2 As shown, the proportion of white corresponding to the instantaneous flash explosion event peaks at the time of the event and decreases within the recovery window after the event; as Figure 3 As shown, the percentage of white corresponding to persistent changes remains offset after the event and does not show a corresponding recovery trend. Therefore, the instantaneous flash explosion event and persistent changes can be distinguished based on the decrease in the percentage of white within the recovery window compared to the peak value of the white percentage event.

[0071] For example, during the recovery mode check, recovery gap protection, extreme flashover protection, and initial gradient protection can be further set. Specifically, when the difference between the peak value of the white percentage event and the minimum white percentage after recovery is greater than 0.005, it is considered that there is a corresponding recovery change after the event, and the event cluster is retained; when the peak value of the white percentage event is greater than 0.50, the event cluster is retained as an extreme flashover case; when the inter-frame difference of the white percentage in the event's starting frame is greater than 0.08, or the inter-frame difference of the overall brightness is greater than 50, it is considered that the initial stage of the event has obvious instantaneous change characteristics, and the event cluster is retained. Event clusters that meet the above protection conditions are not excluded as persistent offset events.

[0072] Example 9:

[0073] This embodiment, based on Embodiment 1, further designs the following: the gradient offset supplementary detection includes pre-checking the change in white proportion between adjacent video frames, performing baseline offset detection based on the median difference in white proportion within the window before and after the video frame, scanning forward when baseline offset is detected, and determining the first video frame that meets the preset offset start point determination condition as the offset start point, verifying the initial change based on the inter-frame change in white proportion within the window before and after the offset start point, verifying the offset persistence based on the situation where the white proportion after the offset start point is continuously higher than the baseline before the offset, and performing color evidence verification. After all verifications are passed, the flash explosion start frame is determined based on the offset start point, and the video frames corresponding to the flash explosion start frame and a preset number of frames before and after it are marked as abnormal frames.

[0074] Specifically, the pre-check includes calculating the maximum difference in white percentage between adjacent video frames. When the maximum value exceeds a preset maximum inter-frame difference threshold (e.g., 0.005), the video is determined to have a significant single-frame jump, and gradient offset supplementation detection is not performed. Baseline offset detection includes calculating the median white percentage in the first and last windows (e.g., the first 5 frames) and the last 5 frames, respectively, and calculating the difference between the two as the baseline offset. When the baseline offset is less than a preset baseline offset threshold (e.g., 0.0025), it is determined that there is no baseline offset. When a baseline offset is detected, the difference between the median highlight percentage in the last and first windows is further calculated. When the highlight percentage offset is less than a preset highlight offset threshold (e.g., 0.0001), it is determined that there is a lack of overexposure evidence, and gradient offset supplementation detection is not performed. The offset starting point is the first video frame that, when scanning forward sequentially, satisfies the condition that the deviation of the median white percentage in the preset window (e.g., 5 frames) starting from that frame relative to the baseline before the offset exceeds half of the baseline offset. The initial change verification involves calculating the maximum increment of the white percentage in adjacent video frames within the window containing the offset starting point as the initial jump value. If the initial jump value is less than a preset initial jump threshold (e.g., 0.0008), the initial change verification fails. It also compares the initial gradient within the window containing the offset starting point with the gradient within the window before the offset. If the initial gradient is less than a preset multiple of the gradient before the offset (e.g., 1.3 times), the initial change verification fails. The offset persistence verification involves counting the number of frames after the offset starting point where the white percentage is higher than the baseline before the offset and the baseline offset by a preset percentage (e.g., 30%). If this number of frames is less than a preset number of frames (e.g., the larger of 30% of the remaining frames and 3 frames), the offset persistence verification fails. The color evidence verification involves counting the number of video frames where the red percentage exceeds a preset percentage (e.g., 0.3%). If this number of frames is less than a preset number (e.g., 5 frames), the color evidence verification fails. After all verifications are passed, the flash explosion start frame is determined based on the offset start point and the initial change verification. The video frames corresponding to the flash explosion start frame and a preset number of frames before and after it are marked as abnormal frames, and the flash explosion fault detection result is output. For example, the preset number can be set to 0, that is, only the video frames corresponding to the flash explosion start frame are marked as abnormal frames.

[0075] Example 10:

[0076] The present invention provides a video flash explosion fault detection system for transmission lines based on dynamic baseline multidimensional feature scoring, comprising a video input module, a feature extraction module, a dynamic baseline module, a detection and judgment module, a gradual offset module, and a result output module.

[0077] The video input module is used to acquire monitoring video of power transmission lines;

[0078] The feature extraction module is used to extract multi-dimensional visual features of video frames one by one. The multi-dimensional visual features include overall brightness, white ratio, highlight ratio, texture complexity, and region features.

[0079] The dynamic baseline module is used to establish corresponding dynamic baselines and calculate standard deviations based on the multidimensional visual features of historical video frames.

[0080] The detection and judgment module is used to determine the dynamic threshold based on the preset absolute threshold and standard deviation, determine the abnormal features based on the comparison between the change of the multidimensional visual features of the current frame relative to the corresponding dynamic baseline and the corresponding dynamic threshold, score the abnormal features and determine flash explosion candidate frames based on the comprehensive score, merge temporally continuous flash explosion candidate frames into events, perform duration filtering based on the duration of each event, and perform recovery mode checks on the flash explosion candidate frames retained after duration filtering to exclude continuous changes and obtain frame-level detection results.

[0081] The gradient offset module is used to perform gradient offset supplementary detection on the white proportion sequence of video frames when no abnormal frames are detected in the frame-level detection results. Based on the offset of the white proportion baseline, it determines the flash explosion event that does not produce a significant single-frame jump and the corresponding abnormal frame.

[0082] The results output module is used to output the flashover fault detection results of the transmission line video based on the frame-level detection results or the detection results of the gradient offset module.

[0083] Specifically, in a CUDA-enabled computing environment, image processing operations such as HSV color space conversion, threshold mask generation, grayscale threshold binarization, and Laplacian operation can be sent to the GPU for execution, while the CPU performs regional grid statistics and subsequent detection and judgment processing to form a video processing process that is collaborative between the GPU and the CPU.

[0084] For example, the system can also be set to a standard detection mode and a high-sensitivity detection mode, and the corresponding detection parameters can be set through a configuration file. The high-sensitivity detection mode uses a relatively low detection threshold to detect weak flash explosion events.

[0085] Specifically, the standard detection mode and the high-sensitivity detection mode each set their detection parameters through corresponding configuration files. For example, the inter-frame brightness jump threshold is set to 30 in the standard detection mode and 0.15 in the high-sensitivity detection mode. Both use the absolute difference in overall brightness between adjacent video frames as the judgment criterion. The high-sensitivity detection mode also sets the starting frame jump threshold to 0.15 and further enables gradient offset supplementary detection to detect weak flash explosion events.

[0086] In one exemplary implementation, the system can also receive the video to be analyzed via an HTTP service interface, supporting synchronous and asynchronous analysis. In asynchronous analysis mode, a corresponding task identifier is established for the video to be analyzed, and the analysis status is queried according to the task identifier. After the analysis is completed, the detection result is returned.

[0087] Application Examples:

[0088] This embodiment employs the aforementioned method for detecting flashover faults in transmission line videos based on dynamic baseline multidimensional feature scoring to detect such faults in transmission line monitoring videos. The experimental computing environment used Linux Ubuntu 20.04 or later, the computer vision framework was OpenCV 4.8.0 with CUDA support enabled, and the programming language was Python 3.9 or later. In the performance testing, the GPU used was an NVIDIA GeForce RTX 3060 with 12GB of video memory, the CPU was an Intel i7-12700H, the test video resolution was 1920×1080, the video duration was 60 seconds, and the frame rate was 25fps.

[0089] The experimental procedure is as follows:

[0090] The monitoring video of the power transmission line to be detected is acquired, and the overall brightness, white ratio, highlight ratio, texture complexity, and regional features are extracted frame by frame according to the method described in the above embodiment. A dynamic baseline is established based on historical video frames, and the standard deviation is calculated. Anomaly judgment and scoring of each dimension of visual features are performed based on dynamic thresholds to determine flashover candidate frames. Temporally continuous flashover candidate frames are merged to form events, and duration filtering and recovery mode checks are performed sequentially. When no abnormal frames are detected at the frame level, a gradual offset supplementary detection is performed on the white ratio sequence to finally obtain the flashover fault detection result. In the detection effect test, 43 labeled videos are detected respectively, and the detection results are compared with the corresponding labeled results. In the processing performance test, the test videos are analyzed using pure CPU processing and GPU accelerated processing respectively, and the time for single-frame feature extraction and the total video analysis time are recorded.

[0091] The test results are as follows:

[0092] Based on existing experimental results, when testing 43 labeled videos, 19 out of 22 videos containing flashover events were correctly detected, 3 were not detected, and all 21 videos without flashover events were correctly detected without false alarms. The recall rate per video was approximately 86.4%, the precision was 100%, and the accuracy was approximately 93.0%. In the aforementioned video processing performance test at 1920×1080 resolution, 60 seconds in length, and 25fps, single-frame feature extraction using pure CPU processing took approximately 12.3ms, with a total analysis time of approximately 102.1s; with GPU acceleration, single-frame feature extraction took approximately 4.1ms, with a total analysis time of approximately 6.5s. These experimental results demonstrate that the method of this invention can detect flashover events in power transmission line monitoring videos and can reduce video analysis time through GPU and CPU collaborative processing.

[0093] Specifically, among the 43 labeled videos, 22 contained flash explosion events, and 21 did not (14 of which were normal detection scenes and 7 were scenes prone to false alarms). The detection results were as follows: of the 22 videos containing flash explosion events, 19 were correctly detected and 3 were missed; of the 21 videos not containing flash explosion events, all were correctly detected and 0 were false alarms. The video-level recall rate was approximately 86.4%, precision was 100%, and accuracy was approximately 93.0%. It should be noted that GPU acceleration improves single-frame image processing (including HSV color space conversion, grayscale threshold binarization, Laplacian operations, etc.) by approximately 3 times, corresponding to the acceleration effect of the image processing computation kernel; while the acceleration rate for total video analysis time is approximately 15.7 times, corresponding to the end-to-end analysis time comparison including frame reading, feature extraction, and detection judgment. The statistical methods for these two methods differ, hence the numerical differences.

[0094] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention 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 the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting flash explosion faults in transmission lines based on dynamic baseline multidimensional feature scoring, characterized in that, include: Acquire monitoring videos of power transmission lines and extract multidimensional visual features from each video frame. The multidimensional visual features include overall brightness, white ratio, highlight ratio, texture complexity, and regional features. Based on the multidimensional visual features of historical video frames, corresponding dynamic baselines are established and their standard deviations are calculated. Dynamic thresholds are determined based on preset absolute thresholds and the standard deviations. Abnormal features are determined based on the comparison between the change of the multidimensional visual features of the current frame relative to the corresponding dynamic baseline and the corresponding dynamic threshold. The abnormal features are scored, and flash explosion candidate frames are determined based on the comprehensive scores. Temporally consecutive flash burst candidate frames are merged into events. Duration filtering is performed based on the duration of each event. The flash burst candidate frames retained after duration filtering are then subjected to recovery mode checks to exclude persistent changes, thus obtaining frame-level detection results. When no abnormal frame is detected in the frame-level detection results, a gradual offset supplementary detection is performed on the white proportion sequence of the video frame, and the flash explosion event and the corresponding abnormal frame are determined based on the offset of the white proportion baseline. Based on the frame-level detection results or the results of the gradient offset supplementary detection, the flashover fault detection results of the transmission line video are output.

2. The method for detecting flashover faults in transmission lines based on dynamic baseline multidimensional feature scoring according to claim 1, characterized in that, The overall brightness is obtained by converting the video frame into a grayscale image and calculating the grayscale mean; the white proportion is obtained by converting the video frame to the HSV color space and calculating the proportion of pixels that simultaneously meet preset low saturation and preset high brightness conditions to the total pixels of the video frame; the highlight proportion is obtained by thresholding the grayscale image of the video frame and calculating the proportion of pixels exceeding a preset high brightness threshold to the total pixels of the video frame; the texture complexity is obtained by calculating the Laplacian variance of the grayscale image of the video frame; the multidimensional visual features also include the red proportion, which is obtained by detecting red pixels in the two hue ranges of 0°~10° and 160°~180° in the HSV color space that meet preset saturation and preset brightness conditions, and calculating the proportion of red pixels to the total pixels of the video frame.

3. The method for detecting flashover faults in transmission lines based on dynamic baseline multidimensional feature scoring according to claim 1, characterized in that, The extraction of the regional features includes dividing the video frame into a 4×4 grid, calculating the local brightness, local white ratio, and local highlight ratio in each grid, and determining the maximum white ratio and maximum highlight ratio of the region based on the calculation results corresponding to each grid.

4. The method for detecting flashover faults in transmission lines based on dynamic baseline multidimensional feature scoring according to claim 1, characterized in that, The step of establishing corresponding dynamic baselines and calculating standard deviations based on the multidimensional visual features of historical video frames includes establishing and maintaining a feature history queue with a sliding window of a preset length, calculating the median and standard deviation of each dimension of visual features in the feature history queue, and using the median as the dynamic baseline of the corresponding visual feature; for each dimension of visual features of the current frame, taking the larger value between the corresponding preset absolute threshold and twice the corresponding standard deviation as the dynamic threshold.

5. The method for detecting flashover faults in transmission lines based on dynamic baseline multidimensional feature scoring according to claim 3, characterized in that, When scoring each abnormal feature, 2 points are awarded when the proportion of white suddenly increases beyond the corresponding dynamic threshold, 2 points are awarded when the proportion of highlight suddenly increases beyond the corresponding dynamic threshold, 1 point is awarded when the overall brightness suddenly increases beyond the corresponding dynamic threshold, 1 point is awarded when the texture complexity decreases beyond the preset texture decrease threshold, 2 points are awarded when the growth rate of the largest white proportion in a region relative to the corresponding dynamic baseline exceeds 50%, and 2 points are awarded when the growth rate of the largest highlight proportion in a region relative to the corresponding dynamic baseline exceeds 50%. The preset score threshold for the comprehensive score is 3 points.

6. The method for detecting flash explosion faults in transmission lines based on dynamic baseline multidimensional feature scoring according to claim 2, characterized in that, The current frame is further detected based on the red proportion. When the growth rate of the red proportion of the current frame relative to the dynamic baseline corresponding to the red proportion exceeds 100%, and the overall brightness of the current frame is higher than the dynamic baseline corresponding to the overall brightness or the absolute value of the overall brightness exceeds a preset brightness threshold, the current frame is determined as a flash explosion candidate frame.

7. The method for detecting flashover faults in transmission lines based on dynamic baseline multidimensional feature scoring according to claim 1, characterized in that, The duration filtering includes recording the duration of each event through an event state machine. When the duration of an event exceeds 20 frames, only the flash burst candidate frame detection results corresponding to the first 5 frames of the event are retained. When the number of flash burst candidate frames in an event is less than 2 frames and there is a lack of color change evidence, all flash burst candidate frame detection results in the event are cleared. The color change evidence includes a sudden increase in the white ratio of the flash burst candidate frame accompanied by overexposure evidence, or a sudden increase in the maximum highlight ratio of the flash burst candidate frame area.

8. The method for detecting flashover faults in transmission lines based on dynamic baseline multidimensional feature scoring according to claim 1, characterized in that, The recovery mode check includes grouping the flash explosion candidate frames retained after duration filtering into event clusters according to the condition that the frame interval between adjacent flash explosion candidate frames is no more than 2 frames, determining the white percentage event peak of each event cluster, obtaining the white percentage within the recovery window after each event cluster, judging whether the corresponding event cluster has recovery characteristics after instantaneous flash explosion based on the decrease between the white percentage event peak and the white percentage within the recovery window, and excluding event clusters whose white percentage maintains a continuous shift and does not show the recovery characteristics.

9. The method for detecting flashover faults in transmission lines based on dynamic baseline multidimensional feature scoring according to claim 1, characterized in that, The gradient offset supplementation detection includes pre-checking the change in white proportion between adjacent video frames, performing baseline offset detection based on the median difference in white proportion within the window before and after the video, scanning forward video frames when baseline offset is detected, and determining the first video frame that meets the preset offset start point determination condition as the offset start point, verifying the initial change based on the inter-frame change in white proportion within the window before and after the offset start point, verifying the offset persistence based on the situation where the white proportion after the offset start point is continuously higher than the baseline before the offset, and performing color evidence verification. After all verifications are passed, the flash explosion start frame is determined based on the offset start point, and the video frames corresponding to the flash explosion start frame and a preset number of frames before and after it are marked as abnormal frames.

10. A transmission line video flash explosion fault detection system based on dynamic baseline multidimensional feature scoring, characterized in that, It includes a video input module, a feature extraction module, a dynamic baseline module, a detection and judgment module, a gradient offset module, and a result output module; The video input module is used to acquire monitoring video of the power transmission line; The feature extraction module is used to extract multi-dimensional visual features of video frames frame by frame. The multi-dimensional visual features include overall brightness, white ratio, highlight ratio, texture complexity, and region features. The dynamic baseline module is used to establish corresponding dynamic baselines and calculate standard deviations based on the multidimensional visual features of historical video frames. The detection and judgment module is used to determine the dynamic threshold based on the preset absolute threshold and the standard deviation, determine the abnormal features based on the comparison result of the change of the multidimensional visual features of the current frame relative to the corresponding dynamic baseline and the corresponding dynamic threshold, score the abnormal features and determine flash explosion candidate frames based on the comprehensive score, merge the temporally continuous flash explosion candidate frames into events, perform duration filtering based on the duration of each event, and perform recovery mode check on the flash explosion candidate frames retained after duration filtering to exclude continuous changes and obtain frame-level detection results. The gradient offset module is used to perform gradient offset supplementary detection on the white proportion sequence of the video frame when the frame-level detection result does not detect abnormal frames, and to determine the flash explosion event and the corresponding abnormal frame based on the offset of the white proportion baseline. The result output module is used to output the flashover fault detection result of the transmission line video based on the frame-level detection result or the detection result of the gradient offset module.

Images