Power distribution network hot-line work discrimination method based on image identification

Abstract

The invention relates to the technical field of power distribution networks, and particularly discloses a power distribution network hot-line work discrimination method based on image recognition. The method is used for solving the problems that a small target of a grounding ring is difficult to stably identify under the condition of scale and posture change due to factors such as strong light glare, appliance shielding, strong reflection metal highlight and perspective distortion in live-line work of a power distribution network, installation in-place and safe distance are difficult to accurately judge, threshold calibers are not uniform, and linkage lag is caused by time sequence jitter. Comprising the steps of scene acquisition and adaptive homomorphic preprocessing, multi-scale target detection modeling, loss calculation and training convergence, target output and label merging, geometric measurement construction, operation state judgment, safety linkage and back-off alarm, record uploading and quality closed loop. According to the method, through cooperation of the adaptive homomorphic preprocessing, the improved YOLOX embedded with geometric priori, the assembly geometric quantity and the time sequence consistent criterion, the identification and the positioning of the grounding ring under the hard light shielding are more stable and accurate, and the wrong judgment and the missed judgment are reduced.

Description

Technical Field

[0001] This invention relates to the field of power distribution network technology, and in particular to a method for identifying live-line work in power distribution networks based on image recognition. Background Technology

[0002] In live-line work on power distribution networks, replacing grounding rings is a crucial step in ensuring the grounding integrity of towers and the safety of personnel. Robots are typically launched using insulated lifting equipment, followed by a multi-functional operation module that identifies and locates the grounding ring near the conductor. An electric wrench mechanism then performs bolt removal and installation, and finally, the robot is deactivated and reset. To achieve automatic identification of grounding rings during live-line work, an image recognition system based on labels such as stripped conductors needs to be constructed. In scenes with strong sunlight and a bright background sky, images captured by an airborne camera need to be homomorphically filtered to suppress overexposure and restore image details for easy identification. Then, an improved YOLOX-S sensor is used to detect and classify the grounding ring and its spatial relationship with the conductor. Based on the geometric constraints and temporal changes of the labels, four types of work states are identified: not in place, disassembling, installing, and in place. The assembly quality is checked using rules such as the concentricity of the grounding ring and conductor, the clamping closure gap, and the exposed length of the bolts. However, the above methods are limited by factors such as high contrast, large target size range, and variable obstruction and posture in typical power environments, strong light causing high brightness and loss of detail, easy distortion of equalization correction, many anchor frame hyperparameters and sample imbalance, and non-overlapping IoU gradients. They are difficult to stably complete the operation judgment of replacing grounding rings under energized conditions, and cannot provide drive data and early warning information for the end effector so as to issue secondary positioning and torque re-check when it is not in place or the pre-tightening is insufficient. If it crosses the boundary and approaches the live body during installation, it should immediately stop and retreat and issue an alarm. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a method for identifying live-line work in power distribution networks based on image recognition. By using adaptive homomorphic preprocessing, improved YOLOX with embedded geometric priors, and synergistic criteria of assembly geometry and timing consistency, the identification and location of grounding rings under strong light obstruction are more stable and accurate, reducing false and missed judgments.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A method for identifying live-line work in a power distribution network based on image recognition, comprising: Step 1, Scene Acquisition and Adaptive Homomorphic Preprocessing: The airborne camera acquires images of the tower operation area and completes intrinsic parameter calibration and distortion correction. Homomorphic filter parameters are adaptively selected to suppress overexposure and restore details. Step 2, Multi-scale target detection modeling: Input the preprocessed image into the improved YOLOX-S at 640×640, adopt the CSPDarknet backbone, three-scale BiFPN fusion and decoupled head, combine Anchor-Free and SimOTA to allocate positive samples, and enable Mosaic and MixUp data augmentation; Step 3, Loss Calculation and Training Convergence: Train the detector until convergence using CIoU as the regression loss and Focal Loss as the confidence loss; Step 4, target output and label merging: The inference stage outputs three types of target boxes: ground_ring, line, and stripped line, along with box confidence and class probability; Step 5, Geometric Measurement Construction: Calculate the comprehensive overlap quality index based on the spatial relationship between the grounding ring and the conductor. Normalized center offset ratio ,in The distance between the centers of the two circumscribed rectangular prediction boxes. The minimum circumscribed diagonal length and aspect ratio angle difference that simultaneously enclose both frames. , , These are the width and height of the prediction box, respectively. , These are the width and height of the reference frame, respectively. Step 6, Job Status Judgment: ≥0.6 and ≤0.1 and ≤0.15 indicates that it is in place. <0.2 or A value >0.4 indicates the bolt is not properly positioned; the remaining values ​​are determined by combining the texture gradient of the bolt area with the closure gap threshold. Determine whether installation or disassembly is in progress; Step 7, Safety Linkage and Retreat Alarm: When it is determined that the device is not in place, is being installed, or is being dismantled, and the minimum safe distance is reached... When this occurs, a shutdown and retreat mechanism is triggered, along with audible and visual alarms and a fault code is written. This is the minimum safe distance threshold; Step 8, Record and submit for quality closure: Record the judgment results, , , , , The timestamp record is uploaded to the host computer for secondary positioning and record reset, and a grounding ring replacement operation record and on-chain evidence are generated at the same time.

[0005] As one aspect of the present invention, step 1 employs an adaptive homomorphic preprocessing method based on sky segmentation, main light source direction estimation, and regionalized homomorphic gain scheduling, specifically including: First, in HSV space, using the luminance channel saturation The sky mask is obtained by combining the candidate horizontal lines with RANSAC filtering. ;right Channel calculation of principal direction gradient , The direction angle of the main light source. The statistical analysis was conducted within the grounding ring ROI. The inner edge ellipse, inner diameter, and normal vector were obtained using elliptic least-squares fitting, and the conductor direction was determined by Hough line detection. The ratio of metallic highlights to the backlight ratio was statistically analyzed, and the homomorphic filter parameters were regionalized accordingly. Within the grounding ring-lug-bolt ROI... , , , ,exist The union of the background side neighborhood will Lower And Upward , specifically by The values ​​were obtained through offline calibration, table lookup, and linear interpolation. , , These are high-frequency gain, low-frequency gain, and cutoff frequency, respectively. , , These are the proportion of metallic highlights, the proportion of backlight side, and the inner diameter of the grounding ring, respectively.

[0006] As one aspect of the present invention, in step 1, ROI differential light suppression is achieved by using an instrument mask and component subdivision enhancement: the electric wrench and socket are detected in the RGB space using a shape template and a color threshold to obtain the instrument mask. Geometric constraints are used to divide the grounding ring into an inner edge, a screw lug opening, and a bolt area, and these are defined. , The width of the image in the coordinate domain currently used for calculation; for Intra-pixel adopts weak homomorphism. A strong homomorphism is used for the union of the lug opening and the bolt area. , The inner edge of the grounding ring is subjected to a medium-intensity homomorphic approach. And through morphological closing operations, the filtering boundary is limited to 2 to 3 pixels.

[0007] As one aspect of the invention, in step 2, geometric and attitude priors related to the grounding ring replacement operation are introduced into the Anchor-Free detection head: the ground_ring category is constrained and predicted at three scales of 8, 16, and 32. , All belong to Aspect Ratio belong Center sampling radius , Add binary attributes to the classification branch. The label is determined by estimating the angle between the orientation of the grounding ring screw and the direction of the conductor. If the absolute value of the angle does not exceed 20°, the screw is determined to be facing the conductor and is in the installation position, and the binary attribute of the sample is marked as 1. If the absolute value of the angle exceeds 20°, the screw is determined to be deviating from the conductor and is in the detached position, and the binary attribute of the sample is marked as 0.

[0008] As one aspect of the present invention, in step 2, data augmentation employs a combination strategy of circularity-preserving affine mapping, instrument occlusion, glare synthesis, motion blur, and randomization of the metallic appearance domain. Specifically, the rotation angle belongs to... Scene scale belongs to The shear is less than or equal to 0.05, and the roundness error is constrained to be less than or equal to 0.03 by fitting an ellipse; the occlusion is made by cutting and masking with a slender rectangle and an electric wrench-shaped template, so that the occlusion area accounts for about 10% to 35% of the outer frame of the ring; the glare generated by gamma enhancement and radial halo kernel is superimposed on the brightness channel; slight anisotropic motion blur is applied in random directions; and the metal texture of the lugs and the ring surface is randomly replaced to simulate the appearance difference between the old ring and the new ring, with a replacement ratio of 0.2 to 0.5.

[0009] As one aspect of the present invention, in step 3, three continuous attribute regression branches are added to the decoupling detection head: lug closure gap, bolt exposed length, and lug-to-conductor angle. A smoothed absolute error loss is used, and weighted gating is applied based on the grounding ring operation stage. During installation, the weights for lug closure gap and exposed length are 1, and the weight for lug-to-conductor angle is 0.5. During disassembly, the weight for lug-to-conductor angle is 1.0, and the weights for lug closure gap and bolt exposed length are 0.8 and 0.6, respectively. In the in-place stage, the weights for lug closure gap, bolt exposed length, and lug-to-conductor angle are 0.6, 0.8, and 1.0, respectively. The classification end uses a hard-to-modulate loss and assigns a grounding ring category. Weights; during training, homomorphically preprocessed samples are mixed with the original samples at a 1:1 ratio, and the intensity of the synthesized glare is gradually increased according to the course learning strategy.

[0010] As one aspect of the present invention, in step 4, the detection results of the three scales are weighted and fused during the inference phase, and the merging threshold is appropriately reduced to improve recall when the grounding ring work equipment is detected. Kalman smoothing is used to suppress jitter for the temporal attributes of the lug closure gap, bolt exposed length, and lug relative to the conductor angle. Majority voting is performed and monotonicity constraints are applied within a continuous frame window of fixed length 5, requiring the gap and exposed length to continuously decrease during the installation phase. The state update of the in-place status is triggered only when the condition is met.

[0011] As one aspect of the present invention, in step 5, based on the comprehensive overlap quality, center offset ratio and aspect ratio angle difference, assembly quality geometric quantities are introduced as the basis for assembly quality acceptance, including the minimum tangential distance from the inner edge to the conductor, the vertical distance from the ring center to the conductor, the angle between the inner edge normal and the conductor direction, the lug closure gap and the bolt protrusion length. All thresholds are adaptively set according to the inner diameter radius of the ring and the pixel diameter of the conductor, and are obtained through step 1. Finally, the upper limit or interval conditions obtained are used to determine whether the assembly passes or fails.

[0012] As one aspect of the present invention, before making a determination in step 6, the vanishing point is estimated based on the directional relationship of multiple conductors and the plane homography transformation matrix is ​​obtained. After normalizing the image to an approximate frontal viewpoint, the overall overlap quality, center offset ratio, aspect ratio angle difference, the closest fit between the inner edge of the grounding ring and the conductor, the vertical distance from the center of the ring to the conductor, the angle between the inner edge normal and the conductor direction, the minimum gap between the mating end faces on both sides of the screw, and the exposed length of the bolt are recalculated. The threshold is then adaptively fine-tuned based on the normalized domain sharpness and scene scale.

[0013] As one aspect of the present invention, in step 7, the safety linkage and recording mechanism adopts dynamic safety boundaries and event-based evidence storage: the minimum safety distance threshold is calculated based on the rated phase voltage of the line, the design coefficient of the distance along the line, and the conductor swing angle. , , These are the voltage-distance conversion factor and the surface distance correction factor, respectively. Calibrated by insulation testing, Offline calibration is performed based on the conductor's pollution level, humidity, and altitude. The nominal voltage of the circuit being worked on. This represents the minimum surface distance requirement under current environmental and equipment conditions. To convert the dynamic margin introduced by the conductor swing angle into a proportional coefficient for the additional distance, It is the angle between the direction vectors of the conductors in adjacent frames, and automatically increases the safety margin when it detects that the lug is not closed, the closure gap is too large, the bolt is not exposed enough, or the normal angle is too large.

[0014] The technical advantages of the image recognition-based live-line working identification method for power distribution networks of this invention are as follows: This invention achieves stable identification and accurate positioning of grounding rings and wires under strong light, backlight, glare, and appliance obstruction environments through adaptive homomorphic preprocessing, an improved single-stage detection network incorporating geometric and orientation priors, a triple criterion consisting of comprehensive overlap quality, center offset ratio, and aspect ratio angle difference, an acceptance mechanism combining assembly geometry, and temporal consistency constraints through multi-scale result fusion, Kalman smoothing, and consecutive frame majority voting. Furthermore, it incorporates viewpoint normalization based on homography transformation and adaptive threshold tuning based on ring radius and wire pixel diameter. This results in regionalized homomorphic... Gain scheduling enhances metallic details in the ring, lug, and bolt sub-regions while suppressing overflow on the sky and background sides. Circularity preservation affine mapping, instrument occlusion and glare synthesis, motion blur, and randomization of the metallic appearance domain improve the separability and generalization ability of small targets and attitude changes. Triple criteria and joint thresholds with assembly geometry reduce missed and false judgments in the already in-place assessment. Dynamic safety boundaries are adjusted in real time based on line voltage, surface distance, and conductor swing, driving rapid retreat and event-based evidence storage, thus forming an interpretable and traceable quality closed loop, facilitating engineering implementation and cross-site migration and replication. Attached Figure Description

[0015] Figure 1 This is a flowchart of a live-line working identification method for power distribution networks based on image recognition, according to the present invention. Detailed Implementation

[0016] The technical solutions of this invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described technical solutions are only a part of this invention, not all of it. Based on the content of this invention, all other solutions obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0017] like Figure 1 As shown, the present invention proposes a method for determining live-line work in power distribution networks based on image recognition, comprising: Step 1, Scene Acquisition and Adaptive Homomorphic Preprocessing: The airborne camera acquires images of the tower operation area and completes intrinsic parameter calibration and distortion correction. Homomorphic filter parameters are adaptively selected to suppress overexposure and restore details. Step 2, Multi-scale target detection modeling: Input the preprocessed image into the improved YOLOX-S at 640×640, adopt the CSPDarknet backbone, three-scale BiFPN fusion and decoupled head, combine Anchor-Free and SimOTA to allocate positive samples, and enable Mosaic and MixUp data augmentation; Step 3, Loss Calculation and Training Convergence: Train the detector until convergence using CIoU as the regression loss and Focal Loss as the confidence loss; Step 4, target output and label merging: The inference stage outputs three types of target boxes: ground_ring, line, and stripped line, along with box confidence and class probability; Step 5, Geometric Measurement Construction: Calculate the comprehensive overlap quality index based on the spatial relationship between the grounding ring and the conductor. Normalized center offset ratio ,in The distance between the centers of the two circumscribed rectangular prediction boxes. The minimum circumscribed diagonal length and aspect ratio angle difference that simultaneously enclose both frames. , , These are the width and height of the prediction box, respectively. , These are the width and height of the reference frame, respectively. Step 6, Job Status Judgment: ≥0.6 and ≤0.1 and ≤0.15 indicates that it is in place. <0.2 or A value >0.4 indicates the bolt is not properly positioned; the remaining values ​​are determined by combining the texture gradient of the bolt area with the closure gap threshold. Determine whether installation or disassembly is in progress; Step 7, Safety Linkage and Retreat Alarm: When it is determined that the device is not in place, is being installed, or is being dismantled, and the minimum safe distance is reached... When this occurs, a shutdown and retreat mechanism is triggered, along with audible and visual alarms and a fault code is written. This is the minimum safe distance threshold; Step 8, Record and submit for quality closure: Record the judgment results, , , , , The timestamp record is uploaded to the host computer for secondary positioning and record reset, and a grounding ring replacement operation record and on-chain evidence are generated at the same time.

[0018] Compared to existing technologies, this invention proposes a regionalized homomorphic gain scheduling driven by sky segmentation and main light source direction estimation, combined with multi-scale detection using an improved YOLOX-S (CSPDarknet backbone, three-scale BiFPN, decoupled head, anchor-free, and SimOTA), a dimensionless triple criterion based on CIoU / center offset ratio / aspect ratio angle difference, supplemented by stage-based determination of bolt area texture gradient and closure gap, and further enhanced by linked backoff triggered by a minimum safe distance threshold and event-based recording. End-to-end process; overcoming global equalization distortion, scale and attitude ambiguity, and timing jitter-induced errors and misjudgments under conditions of strong light, backlight, glare, and equipment obstruction; thereby achieving stable identification, accurate registration, and traceable quality closed loop for grounding loops, wires, and stripped wires under complex lighting and perspective conditions; and achieving rapid, interpretable, and verifiable in-situ determination and intrinsic safety with a clear threshold. Weighted frame fusion, Kalman smoothing, five-frame majority voting, and monography normalization are introduced to suppress jitter and perspective errors. State determination uses the aforementioned joint threshold, and anomalies are recorded. , , , , The data is timestamped and uploaded to the blockchain for evidence storage. Parameters have clear sources and consistent units, facilitating project implementation and cross-site reuse. Audit reproducibility and traceability are supported.

[0019] It should be noted that step 1 employs an adaptive homomorphic preprocessing method based on sky segmentation, main light source direction estimation, and regionalized homomorphic gain scheduling, specifically including: First, in HSV space, using the luminance channel saturation The sky mask is obtained by combining the candidate horizontal lines with RANSAC filtering. ;right Channel calculation of principal direction gradient , The direction angle of the main light source. The statistical analysis was conducted within the grounding ring ROI. The inner edge ellipse, inner diameter, and normal vector were obtained using elliptic least-squares fitting, and the conductor direction was determined by Hough line detection. The ratio of metallic highlights to the backlight ratio was statistically analyzed, and the homomorphic filter parameters were regionalized accordingly. Within the grounding ring-lug-bolt ROI... , , , ,exist The union of the background side neighborhood will Lower And Upward , specifically by The values ​​were obtained through offline calibration, table lookup, and linear interpolation. , , These are high-frequency gain, low-frequency gain, and cutoff frequency, respectively. , , These are the proportion of metallic highlights, the proportion of backlight side, and the inner diameter of the grounding ring, respectively.

[0020] The sky segmentation, main light source direction estimation, and regionalized homomorphic gain scheduling employed in step 1 are because grounding ring installations often occur in strong light, backlight, and highly reflective metallic scenes. High sky brightness can raise global grayscale, and the mirrored surfaces of the lugs and bolts can produce flares and halos. Traditional global equalization or fixed homomorphic parameters tend to amplify noise and obscure details. By separating the sky region, estimating the main lighting direction, and differentially setting high and low frequency gains and cutoff frequencies within the ring, lugs, bolts, and background zones, overexposure and low-frequency clipping are suppressed while preserving metallic edges and textures, avoiding boundary diffusion and false highlights, and improving grounding ring performance. The signal-to-noise ratio and separability of the inner edge of the ground ring, the opening of the screw lug, and the bolt outline are improved, enhancing the stability of feature extraction and detection regression, reducing false negatives under strong light, and making subsequent geometric calculations and in-place determination more accurate and robust. Meanwhile, the parameters are obtained by linear interpolation from the metal highlight ratio, backlight ratio, and inner diameter through offline calibration and table lookup. This allows for adaptive matching to different environmental sizes and light intensities, reducing the risk of manual adjustment and overfitting. The visibility of the threads and edges is enhanced, facilitating accurate measurement, improving the recall and attitude discrimination accuracy of small targets, and ensuring consistency and traceability across sites and different lenses.

[0021] It should be noted that in step 1, ROI differential light suppression is achieved using tool masks and component subdivision enhancement: electric wrenches and sockets are detected in the RGB space using shape templates and color thresholds to obtain tool masks. Geometric constraints are used to divide the grounding ring into an inner edge, a screw lug opening, and a bolt area, and these are defined. , The width of the image in the coordinate domain currently used for calculation; for Intra-pixel adopts weak homomorphism. A strong homomorphism is used for the union of the lug opening and the bolt area. , The inner edge of the grounding ring is subjected to a medium-intensity homomorphic approach. And through morphological closing operations, the filtering boundary is limited to 2 to 3 pixels.

[0022] In step 1, a differentiated homomorphic light suppression method based on "instrument mask + component subdivision" is used for ROI, and the RGB shape template and color threshold detection are used to obtain... Apply weak homomorphism to the region Apply reinforced isomorphic coating to the lug opening and bolt area. Apply moderate intensity to the inner edge of the grounding ring Furthermore, morphological closing operations are used to limit the filter boundary diffusion to 2-3 pixels. The constraint scale is because electric wrenches / sleeves have strong specular highlights and rapid movement. A uniform gain would amplify false textures and cause halo overflow and edge distortion. High and low frequency gains and cutoff frequencies are configured separately according to functional zones. This not only suppresses the highlights of the equipment and background and avoids cross-zone crosstalk and ringing, but also strengthens the lug gaps and bolt contours and maintains the inner edge roundness and true brightness gradient. This improves the signal-to-noise ratio and separability of key boundaries and textures, reduces false detections and missed detections in strong light and occlusion scenes, and improves the accuracy of in-place determination.

[0023] It should be noted that in step 2, geometric and attitude priors related to the grounding ring replacement operation are introduced into the Anchor-Free detection head: the ground_ring category is constrained and predicted at three scales: step size 8, 16, and 32. , All belong to Aspect Ratio belong Center sampling radius , Add binary attributes to the classification branch. The label is determined by estimating the angle between the orientation of the grounding ring screw and the direction of the conductor. If the absolute value of the angle does not exceed 20°, the screw is determined to be facing the conductor and is in the installation position, and the binary attribute of the sample is marked as 1. If the absolute value of the angle exceeds 20°, the screw is determined to be deviating from the conductor and is in the detached position, and the binary attribute of the sample is marked as 0.

[0024] The reason for introducing geometric and attitude priors that match the ground ring replacement operation in the Anchor-Free sensor head (constraining the ground_ring at three scales of 8 / 16 / 32 steps) is... , All belong to Aspect Ratio belong Center sampling radius , And add a binary attribute to the direction of the screw lug towards the wire in the classification branch. The rule of marking 1 if the angle between the screw lug normal and the conductor direction is ≤20° and 0 otherwise is used because the grounding ring is a near-circular small target in the field, and its scale spans and viewpoint distortion coexist. Unified anchor-free methods without prior knowledge are prone to anchor point drift, chaotic positive and negative sample allocation, and attitude ambiguity. By shrinking the candidate box search space and the center sampling radius, and explicitly encoding the key semantics of "screw lug facing conductor", false matching and regression jitter are reduced, the recall and localization accuracy of small targets are improved, convergence is accelerated, and subsequent... , , , The stage determination provides stable and interpretable feature support.

[0025] It should be noted that in step 2, the data augmentation employs a combination strategy of circular-preserving affine mapping, instrument occlusion, glare synthesis, motion blur, and randomization of the metallic appearance domain. Specifically, the rotation angle belongs to... Scene scale belongs to The shear is less than or equal to 0.05, and the roundness error is constrained to be less than or equal to 0.03 by fitting an ellipse; the occlusion is made by cutting and masking with a slender rectangle and an electric wrench-shaped template, so that the occlusion area accounts for about 10% to 35% of the outer frame of the ring; the glare generated by gamma enhancement and radial halo kernel is superimposed on the brightness channel; slight anisotropic motion blur is applied in random directions; and the metal texture of the lugs and the ring surface is randomly replaced to simulate the appearance difference between the old ring and the new ring, with a replacement ratio of 0.2 to 0.5.

[0026] By combining circularity-preserving affine transformation (rotation −30°~30°, scale 0.6~1.4, shear ≤0.05, and elliptic fitting with a circularity error ≤0.03) with data augmentation, tool occlusion (occlusion area of ​​10%~35% by slender rectangles and electric wrench templates), glare synthesis (gamma enhancement and radial halo kernel superimposed on the brightness channel), slight anisotropic motion blur, and randomization of the metallic appearance domain (replacement ratio of 0.2~0.5 for lugs and toroidal textures), we can systematically simulate real interferences such as strong light, backlight, jitter, occlusion, and material aging. This makes the distribution of grounding ring targets in terms of scale, attitude, and reflection conditions closer to the actual situation, significantly improving the recall of small targets, the separability of boundaries and textures, reducing overfitting, and providing robustness. , , , Equal measurement reduces false positives and false negatives and enhances cross-site generalization and threshold transfer consistency. At the same time, affine circle preservation limits roundness error, maintains inner edge geometric consistency, and avoids deformation from destroying the ring prior. Occlusion and motion blur allow the network to learn tolerance to soft occlusion and wind-induced vibration. Glare synthesis approximates specular highlights and spots, suppressing overexposure pseudo-peaks. Texture randomization crosses the differences between new and old parts and electroplating, weakens appearance domain offset, makes the positive and negative sample boundaries clearer and the gradient more stable, accelerates convergence and improves mAP, precision and recall, and maintains the stability and transferability of criteria and measurement thresholds under different lenses and resolutions.

[0027] It should be noted that in step 3, three continuous attribute regression branches are added to the decoupling detection head: lug closure gap, bolt exposed length, and lug-to-conductor angle. A smoothed absolute error loss is used, and weighted gating is applied based on the grounding ring operation stage. During installation, the weights for lug closure gap and exposed length are 1, and the weight for lug-to-conductor angle is 0.5. During disassembly, the weight for lug-to-conductor angle is 1.0, and the weights for lug closure gap and bolt exposed length are 0.8 and 0.6, respectively. In the in-place stage, the weights for lug closure gap, bolt exposed length, and lug-to-conductor angle are 0.6, 0.8, and 1.0, respectively. The classification end uses a hard-to-modulate loss and assigns a category to the grounding ring. Weights; during training, homomorphically preprocessed samples are mixed with the original samples at a 1:1 ratio, and the intensity of the synthesized glare is gradually increased according to the course learning strategy.

[0028] By adding three continuous attribute regression branches—"lug closure gap, bolt protrusion length, and lug relative to wire angle"—to the decoupled detection head and employing smoothed absolute error, the model explicitly learns key assembly geometry while locating the bounding box. Combined with weighted gating according to the "installation, disassembly, and in-place" stages, the supervisory signal is aligned with the actual operation process, thereby enhancing sensitivity to closure trends and attitude changes, and suppressing irrelevant disturbances and gradient noise. The classification end uses hard-example modulation and weighted grounding loop categories to alleviate positive / negative and inter-class imbalances and stabilize the confidence distribution. During training, homomorphically preprocessed samples are mixed one-to-one with the original samples, and glare intensity is gradually increased according to the course, improving generalization and anti-shake capabilities in strong light and reflection scenarios. The combined effect results in faster convergence, more stable regression, and more interpretable threshold judgments, significantly reducing false positives and false negatives and improving the accuracy and cross-scene robustness of in-place judgments.

[0029] It should be noted that in step 4, the detection results of the three scales are weighted and fused during the inference phase, and the merging threshold is appropriately reduced to improve recall when the grounding ring work equipment is detected. Kalman smoothing is used to suppress jitter for the temporal attributes of the lug closure gap, bolt exposed length, and lug relative to the conductor angle. Majority voting is performed and monotonicity constraints are applied within a continuous frame window of fixed length 5, requiring the gap and exposed length to continuously decrease during the installation phase. The state update of the in-place status is triggered only when the condition is met.

[0030] By performing weighted box fusion on the three-scale detection results during the inference phase, and adaptively lowering the merging threshold when the grounding ring work equipment is detected to improve recall, Kalman smoothing is used to suppress jitter for the temporal quantities of lug closure gap, bolt exposed length and lug relative to conductor. Then, majority voting is performed within a 5-frame sliding window and the constraint that "the gap and exposed length during the installation phase must decrease monotonically" is applied. The "in place" state update is triggered only when the monotonicity and stability conditions are met, thereby significantly reducing instantaneous false detections and state repetitions, and improving the consistency, robustness and traceability of positioning and judgment.

[0031] It should be noted that in step 5, based on the comprehensive overlap quality, center offset ratio and aspect ratio angle difference, assembly quality geometric quantities are introduced as the basis for assembly quality acceptance, including the minimum tangential distance from the inner edge to the conductor, the vertical distance from the ring center to the conductor, the angle between the inner edge normal and the conductor direction, the lug closure gap and the bolt protrusion length. All thresholds are adaptively set based on the inner diameter radius of the ring and the pixel diameter of the conductor, and are obtained through step 1. Finally, the upper limit or interval conditions obtained are used to determine whether it passes or fails.

[0032] By introducing assembly quality geometric quantities centered on fitting residuals, center offset, normal angle, lug closure gap, and bolt protrusion length, based on comprehensive overlap quality, center offset ratio, and aspect ratio angle difference, this method adapts and sets thresholds for these quantities based on the inner diameter radius of the ring and the pixel diameter of the wire, derived from scale estimation during the preprocessing stage. This expands the determination of "in place / out of place / in installation / disassembly" from simple frame-level similarity to a structural-level metric, taking into account position, orientation, and assembly completion. This significantly improves the stability and interpretability of the criteria under strong light, backlight, glare, and occlusion conditions, and reduces false detections. The system supports consistent acceptance criteria across lenses and multiple sites, facilitating closed-loop quality inspection and audit reproduction. It also normalizes various geometric quantities and radii and outputs them in conjunction with thresholds to achieve real-time acceptance and anomaly location, reducing repeated manual parameter tuning and accelerating model deployment. In addition, the combination of geometric quantities and phased monotonic constraints can continuously track gap convergence and exposure changes during the installation process, providing early warnings for potential problems such as eccentricity, tilt, and abnormal fit. Thresholds and timestamps are recorded to facilitate on-chain evidence storage and verification. Furthermore, the system ensures that the viewing angle and resolution are on the same scale through single-view normalization, forming a more transferable, quantifiable, and reusable quality standard.

[0033] It should be noted that before the judgment in step 6, the vanishing point is estimated based on the directional relationship of multiple conductors and the plane homography transformation matrix is ​​obtained. After the image is normalized to an approximate frontal view, the overall overlap quality, center offset ratio, aspect ratio angle difference, the closest fit between the inner edge of the grounding ring and the conductor, the vertical distance from the center of the ring to the conductor, the angle between the inner edge normal and the conductor direction, the minimum gap between the mating end faces on both sides of the screw, and the exposed length of the bolt are recalculated. The threshold is then adaptively fine-tuned based on the normalized domain sharpness and scene scale.

[0034] By estimating the vanishing point based on the trends of multiple traverses and obtaining the planar homography transformation before judgment, the image is uniformly normalized to an approximate frontal viewpoint. Then, the geometric quantities of comprehensive overlap quality, center offset ratio, aspect ratio angular difference, as well as fitting residual, center vertical distance, normal angle, lug closure gap, and bolt exposed length are recalculated. The threshold is adaptively fine-tuned based on the normalized domain sharpness and scene scale to eliminate perspective distortion and scale differences, improve measurement consistency and threshold transferability under different lenses, angles, and distances, reduce false detections, missed detections, and jitter caused by viewpoint bias and image blur, and improve the stability, accuracy, and verifiability of the already determined judgment. Simultaneously, the scale and sharpness weights introduced by multi-frame temporal statistics further suppress measurement deviations in highly reflective metals and low-texture areas, ensuring... , , The standardization and audit reproduction of assembly geometry measurements across resolution and site data provide more reliable preliminary measurements for subsequent safety linkage and quality closure.

[0035] It should be noted that in step 7, the safety linkage and recording mechanism adopts dynamic safety boundaries and event-based evidence storage: the minimum safe distance threshold is calculated based on the rated phase voltage of the line, the design coefficient of the distance along the line, and the conductor swing angle. , , These are the voltage-distance conversion factor and the surface distance correction factor, respectively. Calibrated by insulation testing, Offline calibration is performed based on the conductor's pollution level, humidity, and altitude. The nominal voltage of the circuit being worked on. This represents the minimum surface distance requirement under current environmental and equipment conditions. To convert the dynamic margin introduced by the conductor swing angle into a proportional coefficient for the additional distance, It is the angle between the direction vectors of the conductors in adjacent frames, and automatically increases the safety margin when it detects that the lug is not closed, the closure gap is too large, the bolt is not exposed enough, or the normal angle is too large.

[0036] By The system constructs a minimum safe distance threshold that adapts in real time to voltage levels, surface requirements, and conductor sway. It maps physical distances to the pixel domain for online comparison via homography or scene scale. When the system determines that the device is not in place, is being installed, or is being disassembled, and the measured minimum distance is below the threshold, it immediately triggers a shutdown and audible / visual alarm. When high-risk signs such as unclosed lugs, excessive closure gaps, insufficient bolt exposure, or excessive normal angles occur, the system automatically increases the safety margin and adds hysteresis and a minimum hold time to prevent threshold jitter from causing repeated start-stop cycles. This allows the safety boundary to adapt to electrical strength, pollution and humidity, altitude, and wind-induced disturbances, significantly reducing the risk of electric shock and flashover while simultaneously lowering false alarm and missed alarm rates. The system also integrates criteria, minimum safe distance thresholds, and other parameters. , , Equivalent to timestamp-based event submission and on-chain evidence storage, forming a traceable and inherently secure closed loop, with clear sources of thresholds and parameters that can be reused across sites, supporting operation and maintenance supervision and compliance auditing.

[0037] In summary, this invention achieves stable identification and accurate positioning of grounding rings and wires under strong light, backlight, glare, and appliance obstruction environments through adaptive homomorphic preprocessing, an improved single-stage detection network incorporating geometric and attitude priors, a triple criterion consisting of comprehensive overlap quality, center offset ratio, and aspect ratio angle difference, an acceptance mechanism combining assembly geometry, and temporal consistency constraints through multi-scale result fusion, Kalman smoothing, and consecutive frame majority voting. Furthermore, it incorporates viewpoint normalization based on homography transformation and adaptive threshold tuning based on ring radius and wire pixel diameter. Homomorphic gain scheduling enhances metallic details in the ring, lug, and bolt sub-regions while suppressing overflow on the sky and background sides. Circularity preservation affine mapping, instrument occlusion and glare synthesis, motion blur, and randomization augmentation of the metallic appearance domain improve the separability and generalization ability of small targets and attitude changes. Triple criteria and joint thresholding of assembly geometry reduce missed and false judgments in the already in-place assessment. Dynamic safety boundaries are adjusted in real time based on line voltage, surface distance, and conductor swing, driving rapid retreat and event-based evidence storage, thus forming an interpretable and traceable quality closed loop, facilitating engineering implementation and cross-site migration and replication.

[0038] 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.

[0039] Finally: The above description is only for specific explanation of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An image recognition-based power distribution network live working discrimination method, characterized in that, Comprise: Step 1, scene acquisition and adaptive homomorphic preprocessing: the airborne camera acquires the tower operation area image and completes the intrinsic parameter calibration and distortion correction, and adaptively selects the homomorphic filter parameters for overexposure suppression and detail recovery; Step 2, multi-scale target detection modeling: improve YOLOX-S by inputting the preprocessed image with a size of 640x640, use CSPDarknet backbone, three-scale BiFPN fusion and decoupled Head, combine Anchor-Free and SimOTA to assign positive samples, and enable Mosaic and MixUp data augmentation; Step 3, loss calculation and training convergence: train the detector with CIoU as the regression loss and Focal Loss as the confidence loss until convergence; Step 4, target output and label merging: output ground_ring, line, and bopi three types of target boxes, box confidence and class probability in the inference stage; Step 5, Geometric metric construction: Calculate the comprehensive overlap quality index according to the spatial relationship between the grounding ring and the wire , normalized center offset ratio , wherein is the center distance of the two circumscribed rectangle prediction boxes, is the minimum circumscribed diagonal length that simultaneously encloses the two boxes, the aspect ratio angle difference , , are the width and height of the prediction box, respectively , are the width and height of the reference box, respectively Step 6, job status decision: with ≥ 0.6 and ≤ 0.1 and ≤ 0.15 is determined to be seated, with < 0.2 or > 0.4 is determined to be unseated, the rest of the bonded bolt area texture gradient and the closed gap threshold is determined to be in installation or disassembly; Step 7, Safety Linkage and Backoff Alarm: When determined to be unseated, installing or uninstalling and minimum safety distance threshold is triggered, stop backoff, audible and visual alarms and write fault code, is minimum safety distance threshold; Step 8, record upload and quality closed loop: upload the discrimination result, , , , , and the timestamp record to the upper computer for secondary positioning and record reset, and generate the ground ring replacement operation record and on-chain storage certificate.

2. The image recognition-based live working discrimination method for power distribution networks according to claim 1, characterized in that, Step 1 uses an adaptive homomorphic preprocessing method that includes sky segmentation, main light source direction estimation, and regionalized homomorphic gain scheduling, which specifically includes: First, in HSV space with the luminance channel , saturation and combined horizon candidate line RANSAC filtering to get the sky mask ; on channel compute the dominant direction gradient , for the dominant light source direction angle, for the region participating in statistics; in the grounding ring ROI, get the inner edge ellipse and the inner diameter, normal vector by ellipse least square fitting, and get the lead direction by Hough line detection; statistics the metal highlight proportion and the backlight side proportion, and accordingly implement regionalized setting for the homomorphic filtering parameters, take , , , in the the union of the neighborhood on the background side down and up , specifically by linear interpolation from the offline calibration lookup table, , , respectively for the high frequency gain, low frequency gain, cutoff frequency, , , respectively for the metal highlight proportion, the backlight side proportion, and the inner diameter of the grounding ring.

3. The image recognition-based live working discrimination method for power distribution networks according to claim 2, characterized in that, In step 1, ROI differentiation suppression light is enhanced by using an instrument mask and part subdivision: the electric wrench and sleeve are detected by shape template and color threshold in RGB space to obtain an instrument mask ; The ground ring is divided into inner edge, screw ear opening, and bolt area by geometric constraints, and the division is limited , is the image width in the coordinate domain used for the current calculation; for the inner pixels, a weak homomorphism is used, for the union of the screw ear opening and the bolt area, a strong homomorphism is used, , for the inner edge of the ground ring, a medium-strength homomorphism is used, and the filtering boundary is limited to spread to 2 to 3 pixels through a morphological closing operation.

4. The image recognition-based live working discrimination method for power distribution networks according to claim 3, characterized in that, In step 2, the Anchor-Free detector head is introduced with geometry and pose priors related to ground ring replacement jobs: ground_ring class constraint bounding boxes are predicted at three scales of step 8, 16, 32, , all belong to , aspect ratio belong to , center sampling radius , ; Increasing binary attributes in the classification branch The wire side of the screw ear is characterized by first estimating the included angle between the orientation of the ground ring screw ear and the direction of the wire, and if the absolute value of the included angle is not more than 20°, it is determined that the screw ear is directly opposite the wire and is in the position to be installed, and the binary attribute of the sample is marked as 1; if the absolute value of the included angle exceeds 20°, it is determined that the screw ear deviates from the wire and is on the side away from the wire, and the binary attribute of the sample is marked as 0.

5. The image recognition-based live working discrimination method for power distribution networks according to claim 1, characterized in that, In step 2, data augmentation adopts a combination strategy of affine preserving, tool occlusion, glare synthesis, motion blur and metal appearance domain randomization, specifically: the rotation angle belongs to The scene scale belongs to , the shear is less than or equal to 0.05, and the fitting error of the ellipse to the roundness is less than or equal to 0.03; the occlusion area is about 10% to 35% of the ring circumscribed frame by cutting and occluding with an elongated rectangular and a power wrench shape template; the glare is superimposed on the brightness channel by gamma enhancement and radial halo kernel generation; a slight anisotropic motion blur is applied in a random direction; and the metal texture of the screw ear and the ring surface is randomly replaced to simulate the appearance difference between the old ring and the new ring, with a replacement ratio of 0.2-0.

5.

6. The image recognition-based live working discrimination method for power distribution networks according to claim 1, characterized in that, In step 3, three continuous attribute regression branches of screw ear closing gap, bolt exposed length and screw ear relative conductor angle are added in the decoupling detection head. Smooth absolute error loss is adopted, and weight gating is performed according to the grounding ring operation stage. In the installation process, the weights of screw ear closing gap and exposed length are 1, and the weight of screw ear relative conductor angle is 0.

5. In the disassembly process, the weight of screw ear relative conductor angle is 1.0, and the weights of screw ear closing gap and bolt exposed length are 0.8 and 0.6 respectively. In the already positioned stage, the weights of screw ear closing gap, bolt exposed length and screw ear relative conductor angle are 0.6, 0.8 and 1.0 respectively. The classification end adopts loss of difficult example modulation, and assigns weights to the grounding ring categories ​ During training, mix homomorphic preprocessed samples with original samples at a ratio of 1:1, and gradually increase the intensity of synthetic glare according to the curriculum learning strategy.

7. The image recognition-based live working discrimination method for power distribution networks according to claim 1, characterized in that, In step 4, the inference stage implements weighted box fusion on the detection results of the three scales, and appropriately reduces the merging threshold when the grounding ring operating tool is detected to improve recall; Kalman smoothing is used to suppress jitter for the time sequence attributes of the screw ear closure gap, bolt exposure length, and screw ear relative to the wire angle; In a continuous frame window of fixed length 5, majority voting is performed and monotonicity constraints are applied, requiring the installation stage gap and exposure length to continuously decrease, and meeting the conditions to trigger the status update of the installed state.

8. The image recognition-based live working discrimination method for power distribution networks according to claim 1, characterized in that, In step 5, based on the comprehensive overlap quality, center offset ratio, and aspect ratio angle difference, introduce the assembly quality geometric quantity as the assembly quality acceptance basis, including the minimum tangential distance from the inner edge to the wire, the vertical distance from the ring center to the wire, the angle between the inner edge normal and the wire direction, the screw ear closure gap and the bolt exposure length. All thresholds are adaptively set based on the inner diameter radius of the ring and the pixel diameter of the wire, and are obtained through step 1. Finally, the upper limit or interval conditions obtained are used to determine pass or fail.

9. The image recognition-based live working discrimination method for power distribution networks according to claim 1, characterized in that, Before step 6, first estimate the vanishing point based on the trend relationship of multiple wires and obtain the plane homography transformation matrix. After normalizing the image to an approximate orthographic view, recalculate the comprehensive overlap quality, center offset ratio, aspect ratio angle difference, and the closest fit between the inner edge of the grounding ring and the wire, the vertical distance from the ring center to the wire, the angle between the inner edge normal and the wire direction, the minimum gap between the two opposite end faces of the screw ear, and the bolt exposure length. Adjust the thresholds adaptively based on the normalized domain clarity and scene scale.

10. The image recognition-based power distribution network live working discrimination method of claim 1, wherein, In step 7, the security linkage and recording mechanism adopts dynamic security boundary and eventized evidence: the minimum safety distance threshold is calculated according to the rated phase voltage of the line, the along-surface distance design coefficient and the conductor swing angle, , , are the voltage distance conversion coefficient and the along-surface distance correction coefficient respectively, calibrated by insulation test, calibrated offline according to the conductor contamination level, wet state and altitude, is the nominal voltage of the operating object circuit, is the minimum along-surface distance requirement under the current environmental and equipment conditions, is the proportional coefficient for converting the dynamic margin introduced by the conductor swing angle into additional distance, is the angle between the adjacent frame conductor direction vectors, and the safety margin is automatically increased when the screw ear is not closed, the closed gap is out of limit, the screw bolt is insufficiently exposed or the normal angle is out of limit.

Images