A transmission line hidden danger target detection test adaptive method and system

Abstract

The present application belongs to the field of power system intelligent inspection and computer vision technology, and provides a kind of transmission line hidden danger target detection test adaptive method and system, obtains inspection image, utilizes pre-training detector to carry out forward inference to the inspection image, generates original prediction set, and the original prediction set includes prediction frame coordinates, class label and confidence score;Geometric verification is carried out, and reasonable geometric prediction set is obtained;Carry out spatial neighborhood clustering, obtain multiple candidate clusters, calculate the scale dispersion of the area of the prediction frame in each candidate cluster, group according to the class label, dynamically cluster the prediction frame in the group, perform weighted average on the prediction frame coordinates with the confidence score as the weight, calculate the fusion confidence, obtain the refined pseudo label set, use it as the supervision signal, calculate the detection loss, and iteratively update the model parameters through the optimization algorithm until the iteration number is satisfied.The present application improves accuracy, versatility and engineering reliability.

Description

Technical Field

[0001] This invention belongs to the field of intelligent inspection technology for power systems, specifically relating to an adaptive method and system for detecting and testing potential hazards on transmission lines. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the advancement of new power system construction and the development of intelligent inspection technology, online monitoring of potential hazards in transmission lines has become an important part of power grid operation and maintenance management. Various types of inspection image data (such as drone aerial images, images from visual monitoring terminals, and substation perimeter images) are constantly being collected. However, these data often originate from heterogeneous imaging equipment, distributed platforms, and complex inspection scenarios. They themselves have large differences in distribution, inconsistent semantics, and inconsistent imaging conditions, posing a huge challenge to hazard identification and analysis.

[0004] Currently, power transmission line hazard detection systems generally face the following typical problems: Significant cross-domain offset makes traditional supervised detection difficult to adapt to actual inspection environments: Transmission lines involve numerous types of hazards, especially given the complex differences in imaging conditions and target representations across different terrains, seasons, weather conditions, and equipment models. Traditional supervised deep learning detection models or fixed-parameter inference methods struggle to cover the cross-terrain, cross-weather, and cross-equipment detection scenarios encountered in actual inspections. Even with detectors trained in the source domain and having fixed parameters after deployment, the average accuracy significantly decreases when the target domain distribution experiences disturbances such as rain and fog obstruction, low nighttime illumination, strong backlighting, and texture blurring.

[0005] Traditional domain adaptation methods are not feasible in engineering: Unsupervised domain adaptation requires continuous access to source domain labeled data in power transmission scenarios where source domain data is kept confidential and target domain data is delivered in a streaming manner, which contradicts the data security requirements of the power industry; Although passive domain adaptation no longer transmits source domain data back, it still assumes that the target domain is available in the form of a complete offline set, which cannot continuously maintain the real-time requirements of online streaming adaptation.

[0006] Lack of pseudo-label quality control for online detection of power transmission hazards: Most current methods lack a systematic pseudo-label quality control mechanism for the highly complex and dynamic imaging scenario of online detection of power transmission hazards. The pseudo-label selection dimension is too single, making it difficult to guarantee the reliability of pseudo-labels. There is no spatial perception verification layer independent of model confidence, resulting in a lack of stable pseudo-label admission anchors in self-training loops or parameter updates. Furthermore, they cannot be universally applicable across different terrains and weather conditions, and the cost of re-tuning parameters is high.

[0007] The detection output relies on a single highest score prediction, lacking robustness and fusion: Current post-processing for detection generally relies on non-maximum suppression or single confidence threshold filtering, lacking the utilization of collective information from redundant predictions, and unable to generate differentiated fusion localization for multiple predictions of the same target; the ability to protect scale consistency is weak, only providing static post-processing, and lacking the ability to suppress scale oscillations under multiple iterations. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes an adaptive method and system for detecting hidden dangers in power transmission lines. This invention employs a pseudo-label quality control mechanism modeled based on the characteristics of power transmission line inspection operations, combined with adaptive self-training loop scheduling capabilities during testing. This enables an automated processing path from raw inspection images to stable hidden danger alarms, improving accuracy, versatility, and engineering reliability.

[0009] According to some embodiments, the present invention adopts the following technical solution: An adaptive method for detecting potential hazards in power transmission lines includes the following steps: Acquire inspection images, perform forward inference on the inspection images using a pre-trained detector, and generate an original prediction set, which includes prediction box coordinates, category labels, and confidence scores. Perform a geometric verification on the original prediction set to obtain a geometrically reasonable prediction set; Spatial neighborhood clustering is performed on the geometrically reasonable prediction set to obtain multiple candidate clusters. The scale dispersion of the prediction box area within each candidate cluster is calculated. For high dispersion clusters where the scale dispersion exceeds the dispersion threshold, only predictions with confidence scores not lower than a set threshold are retained to obtain a scale-consistent prediction set. The scale-consistent prediction set is grouped by category label, and the prediction boxes within the group are dynamically clustered. The prediction box coordinates are weighted by the confidence score, and the fusion confidence score is calculated to obtain a refined pseudo-label set. Using the refined pseudo-label set as a supervision signal, the detection loss is calculated, and the model parameters of the pre-trained detector are iteratively updated through an optimization algorithm until the number of iterations is met. During the update process, a comprehensive quality score is calculated based on the quality index of the refined pseudo-label set. When the comprehensive quality score is lower than a preset threshold, the current round of parameter update is skipped. The updated pre-trained detector is used to perform final inference on the inspection images to obtain the hazard detection results and issue an early warning.

[0010] As an alternative implementation, the process of using a pre-trained detector to perform forward inference on the inspected image to generate an original prediction set includes: the pre-trained detector provides inference services to the outside world through tensor operations; the inference service completes multi-scale feature extraction and candidate box generation based on input parameters, and returns a structured prediction set in a unified format.

[0011] As an alternative implementation, the process of performing geometric verification on the original prediction set to obtain a geometrically reasonable prediction set includes: obtaining a preset geometric threshold, wherein the geometric threshold is a fixed threshold or dynamically determined according to image resolution, camera viewing distance, target category, historical detection scale distribution or line scene configuration; judging according to the geometric threshold; and filtering out the geometrically reasonable prediction set that meets the threshold requirements.

[0012] As a further defined implementation, the geometric threshold includes several of the following: area threshold, aspect ratio threshold, relative area threshold, confidence threshold, and category adaptive threshold.

[0013] As an alternative implementation, the process of spatial neighborhood clustering of the geometrically reasonable prediction set includes: constructing a pairwise intersection-union matrix of prediction boxes, performing connected component clustering on the pairwise intersection-union matrix according to the intersection-union threshold, and clustering prediction boxes with spatial overlap higher than a set value and consistent categories into the same candidate cluster.

[0014] As an alternative implementation, in the process of calculating the scale dispersion of the predicted box area within each candidate cluster, if a candidate cluster is a single-element cluster, then the candidate cluster is directly retained and the calculation process of the scale dispersion of the area is skipped.

[0015] As an alternative implementation, the process of grouping the scale-consistent prediction set by category label, dynamically clustering the prediction boxes within the group, and performing a weighted average of the prediction box coordinates with confidence scores as weights includes: for each prediction in the scale-consistent prediction set, calculating its maximum intersection-union ratio with all existing fusion cluster weighted average boxes; if it exceeds the fusion threshold, merging and updating the weighted average boxes; otherwise, creating a new cluster. The coordinates within a cluster are weighted according to their confidence levels to obtain the coordinates of the fused box. The arithmetic mean of the average confidence level within the cluster and the highest confidence level is used as the fused confidence level. Output the fusion frame coordinates and fusion confidence for each fusion cluster.

[0016] As an alternative implementation, the process of calculating the detection loss and iteratively updating the model parameters of the pre-trained detector through an optimization algorithm includes: obtaining the current frame image and the current model parameters; The pre-trained detector performs forward inference and obtains refined pseudo-labels; The detection loss is calculated using refined pseudo-label geometry as a supervisory signal. The batch normalized parameters and the weights of the detection head are updated using backpropagation and stochastic gradient descent updaters to obtain the updated model parameters. After iterating through the above steps, the final convergence parameters are obtained, and the final inference is performed on the current image frame using the final convergence parameters.

[0017] As an alternative implementation, a comprehensive quality score is calculated based on the quality indicators of the refined pseudo-label set. When the comprehensive quality score is lower than a preset threshold, the process of skipping the current round of parameter updates includes: calculating four types of indicators, including pseudo-label quantity stability, average fusion confidence change, high discrete cluster ratio, and consistency between continuous frame category and location. Calculate the overall quality score according to the preset weights; When the overall quality score reaches the threshold, it will be updated in this round and written into the stable parameters. When the overall quality score is lower than the threshold, depending on the anomaly type, skip the update, reduce the learning rate, tighten the pseudo-label threshold, or fall back to the previous stable parameter. The exception type, context parameters, pseudo-label size, and stability parameters are recorded as feedback samples.

[0018] An adaptive system for detecting potential hazards in power transmission lines includes: The initial inference module is configured to acquire inspection images, perform forward inference on the inspection images using a pre-trained detector, and generate an original prediction set, which includes prediction box coordinates, category labels, and confidence scores. The geometrically perceptive prior filtering module is configured to perform geometric verification on the original prediction set to obtain a geometrically reasonable prediction set; The area consistency constraint module is configured to perform spatial neighborhood clustering on the geometrically reasonable prediction set to obtain multiple candidate clusters, calculate the scale dispersion of the predicted box area within each candidate cluster, and retain only the predictions with confidence scores not lower than a set threshold for high dispersion clusters with scale dispersion exceeding a dispersion threshold to obtain a scale-consistent prediction set. The classification-aware weighted box fusion module is configured to group the scale-consistent prediction set by category label, dynamically cluster the prediction boxes within the group, perform a weighted average of the prediction box coordinates with confidence scores as weights, calculate the fusion confidence, and obtain a refined pseudo-label set. The online parameter update module is configured to use the refined pseudo-label set as a supervision signal to calculate the detection loss, and to iteratively update the model parameters of the pre-trained detector through an optimization algorithm until the number of iterations is met. During the update process, a comprehensive quality score is calculated based on the quality index of the refined pseudo-label set. When the comprehensive quality score is lower than a preset threshold, the current round of parameter update is skipped. The final inference and early warning module is configured to use the updated pre-trained detector to perform final inference on the inspection image, obtain the hidden danger detection result, and issue an early warning.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention uses a geometrically-aware prior filtering mechanism to verify the spatial geometric attributes of the predicted bounding box independently of the model confidence level. Even under the condition of probability calibration failure, it can still prevent geometrically unreasonable predictions from entering the self-training loop from the source, thereby improving the spatial rationality of pseudo-labels and improving detection efficiency and accuracy. This invention quantifies the scale dispersion of multiple predictions of the same target by using an area consistency constraint algorithm based on the coefficient of variation, performs conditional suppression on high dispersion clusters, reduces the entry of gradient signals with contradictory directions into regression loss, avoids performance oscillations and crashes in long sequence adaptation, reduces the amplitude of performance oscillations, and improves the real-time performance and engineering feasibility of online detection. This invention uses a classification-aware weighted bounding box fusion mechanism to aggregate collective positioning information from redundant predictions, correct the random bias of individual predictions, and output refined pseudo-labels that combine high positioning accuracy with robust uncertainty measurement. This invention achieves online passive migration of model parameters from the source domain to the target domain under the condition that the source domain data is inaccessible and the target domain samples arrive in a streaming manner through an adaptive online parameter update mechanism during testing. This improves cross-scenario adaptability, increases detection accuracy, and reduces the probability of false detection. This invention employs a quality monitoring and parameter rollback mechanism to calculate a comprehensive quality score based on the quality index of false labels. When low-quality updates occur, updates are skipped or the system rolls back to a stable parameter snapshot, reducing the continuous impact of accumulated false labels on long sequence detection and ensuring the engineering reliability and stability of the online adaptation process.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1This is a schematic diagram of system deployment in one embodiment; Figure 2 This is a schematic diagram of the system logic module architecture of one embodiment; Figure 3 This is an overall flowchart of a method according to one embodiment; Figure 4 This is a schematic diagram of a spatially-aware pseudo-tag refining pipeline according to one embodiment; Figure 5 This is a flowchart of a geometry-aware prior filter processing embodiment; Figure 6 This is a flowchart illustrating the fusion process of area consistency constraints and classification-aware weighted bounding boxes in one embodiment. Figure 7 This is a flowchart illustrating the adaptive online parameter update during testing, as shown in one embodiment. Figure 8 This is a diagram illustrating a quality monitoring and parameter rollback mechanism in one embodiment. Figure 9 This is a collaborative sequence diagram of power transmission line hazard detection services in one embodiment. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0026] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0027] Example 1 An adaptive method for detecting hidden dangers in transmission lines is proposed. This method is based on an adaptive self-training loop during testing, supported by a spatial-aware pseudo-label refinement pipeline (SRP) and a quality backoff protection mechanism. This method combines geometric admission, scale consistency suppression, classification perception fusion and online parameter protection for the detection of hidden dangers in transmission lines.

[0028] like Figure 3 As shown, it includes the following steps: Step S1: Obtain the inspection image, and use the pre-trained detector to perform forward inference on the inspection image to generate an original prediction set. The original prediction set includes the prediction box coordinates, category labels and confidence scores. Step S2: Perform geometric verification on the original prediction set to obtain a geometrically reasonable prediction set; Step S3: Perform spatial neighborhood clustering on the geometrically reasonable prediction set to obtain multiple candidate clusters. Calculate the scale dispersion of the prediction box area within each candidate cluster. For high dispersion clusters where the scale dispersion exceeds the dispersion threshold, only retain predictions with confidence scores not lower than the set threshold to obtain a scale-consistent prediction set. Step S4: Group the scale-consistent prediction set by category label, perform dynamic clustering on the prediction boxes within the group, perform weighted averaging on the prediction box coordinates with confidence scores as weights, calculate the fusion confidence, and obtain a refined pseudo-label set. Step S5: Using the refined pseudo-label set as a supervision signal, calculate the detection loss, and iteratively update the model parameters of the pre-trained detector through an optimization algorithm until the number of iterations is met. During the update process, calculate the comprehensive quality score based on the quality index of the refined pseudo-label set. When the comprehensive quality score is lower than a preset threshold, skip the current round of parameter update. Step S6: Use the updated pre-trained detector to perform final inference on the inspection image to obtain the hidden danger detection result and issue an early warning.

[0029] The following is a detailed introduction.

[0030] like Figure 1 As shown, in this embodiment, inspection image data (including visible light, infrared illumination, multispectral imaging modalities) can be generated in real time by acquisition devices such as UAV aerial photography systems, fixed visual monitoring terminals, and substation perimeter cameras. These raw images are first entered into the edge image cache and used as the basic input for subsequent inference.

[0031] During the inspection of power transmission lines, the source domain server side collected a large amount of labeled image data of potential hazards (such as bird nests, hanging foreign objects, construction machinery, broken insulators, slippage of anti-vibration hammers, etc.). This data was first used to train a general target detector, i.e., a pre-trained detector.

[0032] During the pre-training phase, supervised learning is performed on the source domain images, modeling the complex hazard identification task as a joint optimization problem of classification and regression, learning the mapping relationship from image pixels to target categories and bounding box coordinates. After pre-training, the model parameters... The data is frozen and sent to the edge inspection device, and the source domain data is no longer transmitted back, thus meeting the engineering constraints of power data confidentiality and edge deployment.

[0033] In this embodiment, the pre-trained detector uses, but is not limited to, the Faster R-CNN detection framework, and the backbone network can be an existing ResNet-50 / ResNet-101, etc. The detector architecture mainly includes: Backbone feature extraction network: maps the input image into a multi-scale feature map; Regional Proposal Network (RPN): Pre-sets anchor boxes of various scales and aspect ratios at each spatial location in the feature map, and outputs foreground confidence and coordinate offset; Detection head: Performs region of interest pooling, category classification, and bounding box fine-tuning on the candidate regions output by RPN; And the detection loss function.

[0034] All of the above are passed in memory in the form of tensors to avoid directly exposing the details of the underlying network and to ensure that the upper part of SRP receives semantically clear structured prediction input.

[0035] Of course, other detectors in the existing technology can also be used for pre-trained detectors, which will not be exhaustively listed here.

[0036] In this embodiment, the pre-trained detector provides inference services through standard tensor operations. The SRP method or an external inspection system can call this inference interface to obtain candidate predictions. The inference service's responsibilities are to receive input parameters, call the detector to complete multi-scale feature extraction and candidate box generation, and return a structured prediction set in a unified format, i.e., the original prediction set. Each prediction contains the coordinates of the prediction box. Category tags With confidence score However, it is not limited to a specific tensor layout form. The number of elements in the set.

[0037] The original prediction set undergoes a three-level refinement process through the SRP pipeline, forming a multi-level pseudo-tag quality control chain. The three levels of refinement are geometric prior verification, statistical analysis, and fusion. By utilizing the progressive filtering effect of the three-level refinement, it is ensured that the pseudo-tag set entering backpropagation meets the quality requirements in terms of geometric rationality, scale consistency, and positioning accuracy.

[0038] In this embodiment, as Figure 5 As shown, a geometric prior verification component performs geometric verification on the original prediction set, independent of the model confidence level. This component receives the prediction set from the pre-trained detector and a geometric threshold parameter; the geometric threshold can be a default threshold or dynamically determined based on image resolution, camera viewing distance, target category, historical detection scale distribution, or line scene configuration, and is converted into a specific judgment operation, outputting a geometrically reasonable prediction set and a record of rejection reasons.

[0039] The geometric prior verification component has the capability to perform geometric determinations. As a pseudo-label admission verification layer, it can proactively prevent geometrically unreasonable predictions from entering the parameter update stage and pass compliant predictions to downstream modules. The components requiring configuration include area threshold, aspect ratio threshold, relative area threshold, confidence threshold, class adaptation threshold, and candidate observation strategy. The parameters passed to this verification node are the predicted bounding box coordinates, class label, confidence score, and total image area.

[0040] Define a validity indicator function: ; The area is calculated for each prediction box. Aspect Ratio With relative coverage ; Small area filtering branches: If , If the area threshold is used, it is identified as rain / fog particles / texture artifacts and removed. Aspect Ratio Extreme Filtering Branch: If , If the aspect ratio threshold is met, then it is identified as a false positive and removed as a thin background such as a wire or ground wire. Large-area low-confidence joint filtering branch: if And confidence level If it is determined to be a large area of ​​background such as the tower / sky that has been mistakenly activated, it will be removed. This is the relative coverage threshold. The confidence score threshold; The intersection of the three rules, executed in parallel, outputs a geometrically reasonable prediction set. ; This is an indicator function that takes the value 1 when the condition within the parentheses is true, and 0 otherwise. It is a logical NOT function. For logical AND.

[0041] The resulting geometrically reasonable prediction set for: ; In this embodiment, as Figure 6 As shown, after receiving the geometrically reasonable prediction set, the area consistency constraint module identifies candidate clusters under the constraints of category consistency and spatial overlap. When the coefficient of variation of the predicted box area within a cluster is lower than the threshold, the cluster is retained and enters the downstream fusion. When the coefficient of variation exceeds the threshold, only predictions that meet the high confidence condition are retained, and the remaining predictions do not participate in this round of online updates. Subsequently, the area consistency constraint module connects the screening results with the downstream fusion module to generate an intermediate prediction set that satisfies the consistency of the regression gradient direction.

[0042] The workflow of the area consistency constraint module is as follows: In the first stage, construct the pairwise intersection-union matrix of the prediction boxes: ; These are the coordinates of the i-th and j-th predicted bounding boxes, respectively.

[0043] With threshold Perform connected component clustering on the pairwise intersection-union matrices to obtain a set of candidate clusters. K represents the total number of candidate clusters.

[0044] In the second stage, the area variation coefficient is calculated for each candidate cluster: ; in, , These are the standard deviation and mean of the predicted box area within the candidate cluster, respectively.

[0045] The third stage, execution condition suppression: ; in, The threshold for the coefficient of variation of area. The confidence score threshold; Final output .

[0046] In this system, single-element clusters are directly retained and the calculation of the area variation coefficient is skipped; the standard deviation calculation method for multi-element clusters is fixed in the system configuration, and either the population standard deviation or the sample standard deviation can be used. To avoid neighboring targets being incorrectly merged, spatial clustering can simultaneously constrain class consistency, centroid distance, or aspect ratio differences.

[0047] After obtaining a scale-consistent prediction set, in this embodiment, as follows: Figure 6 As shown, the Class-Aware Weighted Box Fusion (WBF) module is used to group objects by category and perform dynamic clustering and weighted fusion: For each prediction to be processed, the classification-aware weighted box fusion module calculates the maximum intersection over union (IoU) of the weighted average box with all existing fusion clusters. If it exceeds the fusion threshold, it merges and updates the weighted average box in real time; otherwise, it creates a new cluster. Finally, it outputs the aggregated coordinates and compromise confidence for each fusion cluster.

[0048] This module is based on the input: a scale-consistent prediction set. Fusion threshold The process involves classification and grouping, dynamic clustering, score-weighted coordinate averaging, and fusion confidence estimation to obtain a refined set of pseudo-labels. .in Let be the refined pseudo-label set obtained after processing the t-th frame sample in the τ-th round. Let k be the coordinates of the fused predicted bounding box. Let k be the category label of the fused pseudo-label. is the confidence score of the k-th fusion.

[0049] The key step in generating fused pseudo-labels in this module is to design a suitable aggregation strategy to make full use of the collective positioning information of redundant predictions. In this embodiment, a specified aggregation formula is designed for weighted coordinate fusion and fusion confidence estimation.

[0050] The specific aggregation formula is as follows: ; in, Let k be the coordinates of the fused predicted bounding box. For the k-th fusion cluster, For fusion cluster The confidence score of the i-th predicted bounding box. For fusion cluster The coordinates of the i-th predicted bounding box. The confidence score after fusion is the k-th fusion. For fusion cluster The number of predicted boxes in the middle.

[0051] In addition, the clustering requirements of the classification-aware weighted bounding box fusion module include: The aggregated output should be a standard bounding box coordinate and confidence score tuple, conforming to the input specifications of the detection loss function, with no non-standard fields; All coordinates must be aggregated in a score-weighted manner, and the highest-scoring frame should not be simply selected to maintain consistency with the original predictions. The merged cluster must contain all the necessary elements: the merged prediction index, the weighted average coordinates, the compromise confidence score, and, if necessary, the highest and average scores within the cluster. For clusters containing only a single prediction, the original coordinates and confidence scores should be output directly to avoid unnecessary weighted calculations. These aggregation results are embedded in the output set of pseudo-labels so that they can be used directly in the calculation of self-training loss.

[0052] In this embodiment, refined pseudo-labels serve as monitoring signals, driving online updates of model parameters through backpropagation of detection losses. The final results can either be used to directly render alarms at the edge or uploaded to the main inspection station to generate hazard work orders. This embodiment utilizes online parameter updates and final alarm output modules, supporting deployment at the edge or in a cloud-edge collaborative environment, and can interface with external inspection scheduling platforms via an application programming interface (API).

[0053] like Figure 7 As shown, the input to the online parameter update and final alarm output module is a refined set of pseudo-tags. Current parameters Learning rate The process involves calculating detection loss, backpropagation, parameter updating, iterative convergence, and final inference to obtain deliverable hazard alarm results (structured data exchange format / main station work order / visual annotation).

[0054] Iterate per frame Use final parameters after wheel For images Perform the final inference and output the hazard alarm result.

[0055] To achieve self-optimization and continuous stability in the online detection process of power transmission hazards, this embodiment proposes a multi-round iterative quality monitoring and rollback mechanism, such as... Figure 8 As shown, this mechanism is used for automatic evaluation and adaptive protection of parameter update trajectories in the self-training loop. This mechanism constitutes an adaptive system from image input—pseudo-label refinement—parameter update—quality monitoring—fallback protection.

[0056] The mechanism comprises three core components: First, the quality assessment model is output. The detection results after each iteration are analyzed in a structured manner, and a comprehensive score is calculated from four dimensions: stability of the number of pseudo-labels, trend of average fusion confidence, cluster dispersion distribution, and consistency of detection in consecutive frames.

[0057] Secondly, the feedback optimization strategy. When the evaluation score falls below a set threshold, the system automatically records the anomaly type and context parameters (such as the current iteration round, the size of the pseudo-label set, and the model parameter snapshot) to form an adaptation feedback sample. This sample is used to update the learning rate weights and backoff trigger conditions for subsequent iterations, such as strengthening the conservative update strategy in high-dispersion scenarios and correcting the pseudo-label admission threshold in low-confidence scenarios.

[0058] Finally, multi-round adaptive protection. For low-scoring iteration rounds, the system automatically skips the parameter update for that round or triggers a rollback to the most recent stable snapshot; the new iteration will refer to the correction information in the feedback samples and adjust the update strategy, thereby achieving continuous stability in the adaptation process.

[0059] This feedback mechanism enables the system to possess self-protection capabilities. With the accumulation of inspection tasks, the system's adaptability, stability, detection consistency, and engineering reliability under long-term changing environments gradually improve. This mechanism not only enhances the intelligence level of online detection but also forms a sustainable protection system at the system level, significantly enhancing its engineering practicality.

[0060] Example 2 As a specific application, an adaptive method for detecting hidden dangers in power transmission lines is used, such as... Figure 9 As shown, it includes the following steps: Drones or visual monitoring terminals initiate requests for power transmission line hazard detection, such as "real-time hazard identification of tower YY on the XX power transmission line," and transmit the collected inspection images. Push to the edge inspection device; The edge-end pre-trained detector receives an image and automatically performs forward inference (loading source domain parameters). Or the convergence parameters of the previous frame ), and extract parameters such as candidate prediction box coordinates, category labels, and confidence scores; Perform multidimensional geometric verification on the original prediction set, encapsulate complex geometric decision rules (minimum area, aspect ratio extrema, large area-low confidence joint rejection) and obtain a geometrically reasonable prediction set. ; Based on the geometrically reasonable prediction set, scale consistency verification is performed through spatial neighborhood clustering and coefficient of variation quantization. Conditional suppression is applied to high-dispersion clusters that exceed the threshold, resulting in a scale-consistent prediction set. ; Perform classification grouping, dynamic clustering, and score-weighted coordinate averaging on the scale-consistent prediction set to generate a fused coordinate system. Category tags With compromise confidence level A collection of refined pseudo-tags ; The detection loss is calculated by backpropagation using refined pseudo-tags as the supervision signal, and the batch normalized affine parameters and the weights of the detection head are updated. The above steps are executed in a loop for P rounds, forming a single-frame test-time adaptation loop; Finally, the parameters after the P-round cycle are used. Inspection images The final inference is executed, and the output of the hidden danger detection results can be directly visualized and displayed at the edge, or sent to the inspection master station through links such as 4G / 5G and power private network to form a complete alarm work order delivery.

[0061] Example 3 An adaptive method for detecting potential hazards on power transmission lines, applied to the identification of illegal construction machinery in low-light nighttime environments, includes: Input command: Monitor potential external damage to the XX power transmission channel at night. The trigger condition is a nighttime image captured by the visual monitoring terminal under infrared supplementary lighting.

[0062] Step 1: The edge detector identifies the task as an external damage hazard detection flow and loads the convergence parameters from the previous frame. ; Step 2: Extract parameters: Monitoring point ID=YY-A03, Inspection time=22:30 at night, Imaging mode=infrared supplementary light; Step 3: The pre-trained detector performs forward inference and outputs the original prediction set; Step 4: Intercept predictions of large areas of night sky background that are misclassified as foreground with medium confidence; perform coefficient of variation verification on multiple clusters of excavator predictions, and retain only high-confidence predictions for out-of-bounds clusters; perform weighted fusion on multiple redundant predictions of tower cranes. Step 5: Iteration of the online parameter update module Wheels allow the model to adapt to low-light imaging characteristics at night; Step 6: Finally, the identification frames of the tower crane and excavator are output, triggering the external damage hazard alarm; Step 7: Send the alarm work order to the main station and activate the on-site loudspeaker to warn of violations.

[0063] Example 4 An adaptive system for detecting potential hazards in power transmission lines includes: The initial inference module is configured to acquire inspection images, perform forward inference on the inspection images using a pre-trained detector, and generate an original prediction set, which includes prediction box coordinates, category labels, and confidence scores. The geometrically perceptive prior filtering module is configured to perform geometric verification on the original prediction set to obtain a geometrically reasonable prediction set; The area consistency constraint module is configured to perform spatial neighborhood clustering on the geometrically reasonable prediction set to obtain multiple candidate clusters, calculate the scale dispersion of the predicted box area within each candidate cluster, and retain only the predictions with confidence scores not lower than a set threshold for high dispersion clusters with scale dispersion exceeding a dispersion threshold to obtain a scale-consistent prediction set. The classification-aware weighted box fusion module is configured to group the scale-consistent prediction set by category label, dynamically cluster the prediction boxes within the group, perform a weighted average of the prediction box coordinates with confidence scores as weights, calculate the fusion confidence, and obtain a refined pseudo-label set. The online parameter update module is configured to use the refined pseudo-label set as a supervision signal to calculate the detection loss, and to iteratively update the model parameters of the pre-trained detector through an optimization algorithm until the number of iterations is met. During the update process, a comprehensive quality score is calculated based on the quality index of the refined pseudo-label set. When the comprehensive quality score is lower than a preset threshold, the current round of parameter update is skipped. The final inference and early warning module is configured to use the updated pre-trained detector to perform final inference on the inspection image, obtain the hidden danger detection result, and issue an early warning.

[0064] Example 5 An adaptive system for detecting hidden dangers in power transmission lines consists of three parts: an image acquisition end, an edge inspection device, and an inspection master station. The image acquisition end is used to acquire inspection images captured by drones, fixed visual monitoring terminals, or substation perimeter cameras. The edge inspection device is used to complete image access, pre-trained detector loading, SRP pseudo-label refinement, TTA online adaptation, and alarm encapsulation. The inspection master station is used to receive alarm results, display visual annotations, and generate maintenance work orders.

[0065] Among them, this system is as follows Figure 2As shown, the architecture comprises an input layer, a core processing layer, and an output layer. The input layer receives the inspection image stream, device and task metadata, and scene configuration parameters. The core processing layer includes a pre-trained detector, raw prediction generation, a Geometric Prior Filter (GPF) module, an Area Consistency Constraint (ACC) module based on spatial statistical features, a Class-Aware Weighted Box Fusion (WBF) module, online parameter updates, and quality monitoring and rollback modules. The output layer generates structured alarms, main station interface data, and labeled image results. These modules collaborate through explicit tensor and configuration interfaces, forming a vertical data stream: image input—multi-level pseudo-label refinement—online parameter updates—alarm output.

[0066] The difference from the above embodiments is that, in this embodiment, the processing procedure of the geometry-aware prior filter module includes: calculating the area of ​​each prediction box. Aspect Ratio Relative area; Determine whether small area constraints, aspect ratio extreme value constraints, or large area low confidence constraints are triggered. For prediction records and triggering rules that trigger rejection conditions, exclude them from the pseudo-label set for this round of parameter updates; For predictions that do not trigger the rejection condition, output them to the area consistency constraint module.

[0067] The processing steps of the area consistency constraint module include: performing geometrically reasonable predictions on the set of categories. Perform pre-grouping; Within each category group, pairwise IoU of predicted boxes is calculated, and adjacency relationships can be constructed by combining center point distance or aspect ratio differences; Candidate cluster set is obtained based on adjacency relationship. S4: For single-element clusters, the area is directly retained; for multi-element clusters, the mean area is calculated. Standard deviation and coefficient of variation ; when When all predictions within the cluster are retained, Only retain those that meet the requirements. The remaining predictions will not be included in this round of parameter updates. The threshold for the coefficient of variation of area. This is the confidence score threshold.

[0068] Obtain the scale-consistent prediction set The data includes the member list, area statistics, CV value, decision on whether to retain, and highly discrete label for each candidate cluster.

[0069] The classification-aware weighted bounding box fusion module includes: Classifier grouper: Splits the scale-consistent prediction set into multiple groups based on the category label and sorts them in descending order of confidence; Dynamic clusterer: For each predicted box in a group, calculate the maximum IoU between it and the weighted average box of the existing fused cluster. If the IoU exceeds a preset threshold, merge the predicted boxes and update the weighted average boxes in real time; otherwise, create a new cluster. Fractionally weighted coordinate averager: Calculates the intra-cluster coordinate weighted mean using confidence scores as weights. Compromise confidence estimator: The arithmetic mean of the cluster average and the highest score is used as the fusion confidence; this yields a refined set of pseudo-labels. .

[0070] The specific steps are as follows: Scale-consistent prediction sets according to category labels Group and sort by confidence level in descending order; Dynamic clustering is performed on predicted boxes within the same category. If the weighted average IoU between the box to be processed and an existing fusion cluster is greater than a pre-set fusion IoU threshold, the box is assigned to that cluster; otherwise, a new cluster is created. The coordinates within a cluster are weighted according to confidence level to obtain the coordinates of the fused box. The fusion confidence is the arithmetic mean of the average confidence within the cluster and the highest confidence. For single-element clusters, the original coordinates and confidence scores are directly output.

[0071] In this embodiment, the online parameter update module updates the current frame image with the current model parameters. For input; The detector performs forward inference and invokes the SRP pipeline to generate refined pseudo-labels; Loss detection calculator As a supervisory signal, the detection loss is calculated. ; Backpropagation and the Stochastic Gradient Descent (SGD) updater update the batch-normalized affine parameters and the partial weights of the detector head, resulting in... ; The above steps are iterated Post-round output convergence parameters Perform final inference on the current frame.

[0072] like Figure 8As shown, after each iteration, the quality monitoring and rollback module submits indicators such as pseudo-label set size, average fusion confidence, cluster dispersion distribution, and continuous frame detection consistency to the quality evaluator for item scoring and weighted summation. If the score is higher than the threshold, the current update will be accepted and the parameters will be written to the stable parameter snapshot library. If the score is below the threshold, a suitable feedback sample is constructed by the feedback sample generator. Feedback samples are entered into the policy tuning module to update the learning rate weights, pseudo-label admission thresholds, or backoff trigger conditions for subsequent iterations. When the scores are below the threshold for several consecutive rounds, the system automatically reverts from the stable parameter snapshot library to the most recent stable parameters to avoid the accumulation and amplification of errors in the self-training loop; at the same time, the feedback samples are stored in the experience sample library for long-term strategy optimization.

[0073] The system provided in this embodiment can be deployed in edge inspection devices or cloud-edge collaborative environments. It consists of a pre-trained detector layer, an SRP pipeline layer, an online update layer, an alarm transmission layer, and a master station receiving layer. Each module can be implemented using common deep learning frameworks and edge computing platforms, but the specific implementation language, framework, and hardware platform are engineering details and are not intended to limit the invention.

[0074] like Figure 9 As shown, the timing coordination relationship between the image acquisition terminal, edge inspection device, SRP pipeline, online update module, quality monitoring module, and inspection master station is as follows: The image acquisition terminal is used to: trigger inspection tasks, acquire images, and send them to the edge terminal; The edge inspection device is used for: receiving images - loading parameters - forward inference - SRP refinement - loss calculation - parameter update - quality monitoring - final inference - alarm encapsulation; The main inspection station is used for: receiving alarm work orders, visual display, work order workflow and manual review; The order of input / output is as follows: Inspection task - Image acquisition - Detection inference - Pseudo-label refinement - Parameter update - Final alarm - Main station delivery.

[0075] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0076] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0077] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0078] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0079] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A self-adaptive method for transmission line hazard target detection test, characterized in that, Includes the following steps: Acquire inspection images, perform forward inference on the inspection images using a pre-trained detector, and generate an original prediction set, which includes prediction box coordinates, category labels, and confidence scores. Perform a geometric verification on the original prediction set to obtain a geometrically reasonable prediction set; Spatial neighborhood clustering is performed on the geometrically reasonable prediction set to obtain multiple candidate clusters. The scale dispersion of the prediction box area within each candidate cluster is calculated. For high dispersion clusters where the scale dispersion exceeds the dispersion threshold, only predictions with confidence scores not lower than a set threshold are retained to obtain a scale-consistent prediction set. The scale-consistent prediction set is grouped by category label, and the prediction boxes within the group are dynamically clustered. The prediction box coordinates are weighted by the confidence score, and the fusion confidence score is calculated to obtain a refined pseudo-label set. Using the refined pseudo-label set as a supervision signal, the detection loss is calculated, and the model parameters of the pre-trained detector are iteratively updated through an optimization algorithm until the number of iterations is met. During the update process, a comprehensive quality score is calculated based on the quality index of the refined pseudo-label set. When the comprehensive quality score is lower than a preset threshold, the current round of parameter update is skipped. The updated pre-trained detector is used to perform final inference on the inspection images to obtain the hazard detection results and issue an early warning.

2. A method according to claim 1, wherein the method is further characterized by: The process of using a pre-trained detector to perform forward inference on the inspected image and generate an original prediction set includes: the pre-trained detector provides inference services to the outside world through tensor operations; the inference service completes multi-scale feature extraction and candidate box generation based on input parameters, and returns a structured prediction set in a unified format.

3. The adaptive method for detecting hidden dangers in transmission lines as described in claim 1, characterized in that, The process of performing geometric verification on the original prediction set to obtain a geometrically reasonable prediction set includes: obtaining a preset geometric threshold, wherein the geometric threshold is a fixed threshold or dynamically determined according to image resolution, camera viewing distance, target category, historical detection scale distribution or line scene configuration; judging according to the geometric threshold; and filtering out the geometrically reasonable prediction set that meets the threshold requirements.

4. The adaptive method for detecting hidden dangers in transmission lines as described in claim 3, characterized in that, The geometric thresholds include several of the following: area threshold, aspect ratio threshold, relative area threshold, confidence threshold, and category adaptive threshold.

5. The adaptive method for detecting hidden dangers in transmission lines as described in claim 1, characterized in that, The process of spatial neighborhood clustering of the geometrically reasonable prediction set includes: constructing a pairwise intersection-union matrix of prediction boxes, performing connected component clustering on the pairwise intersection-union matrix according to the intersection-union threshold, and clustering prediction boxes with spatial overlap higher than a set value and consistent categories into the same candidate cluster.

6. A method according to claim 1, wherein the method further comprises: determining the location of the object based on the received signal strength of the signal received from the object. In the process of calculating the scale dispersion of the predicted box area within each candidate cluster, if a candidate cluster is a single-element cluster, the candidate cluster is directly retained and the calculation process of the scale dispersion of the area is skipped.

7. A method according to claim 1, wherein the method is further characterized by: The process of grouping the scale-consistent prediction set by category label, dynamically clustering the prediction boxes within the group, and performing a weighted average of the prediction box coordinates with confidence scores as weights includes: for each prediction in the scale-consistent prediction set, calculating its maximum intersection-union ratio with all existing fusion cluster weighted average boxes; if it exceeds the fusion threshold, merging and updating the weighted average boxes; otherwise, creating a new cluster. The coordinates within a cluster are weighted according to their confidence levels to obtain the coordinates of the fused box. The arithmetic mean of the average confidence level within the cluster and the highest confidence level is used as the fused confidence level. Output the fusion frame coordinates and fusion confidence for each fusion cluster.

8. A method according to claim 1, wherein the method is further characterized by: The process of calculating the detection loss and iteratively updating the model parameters of the pre-trained detector through an optimization algorithm includes: obtaining the current frame image and the current model parameters; The pre-trained detector performs forward inference and obtains refined pseudo-labels; The detection loss is calculated using refined pseudo-label geometry as a supervisory signal. The batch normalized parameters and the weights of the detection head are updated using backpropagation and stochastic gradient descent updaters to obtain the updated model parameters. After iterating through the above steps, the final convergence parameters are obtained, and the final inference is performed on the current image frame using the final convergence parameters.

9. A method according to claim 1, wherein the method is further characterized by: The process of skipping the parameter update in this round when the comprehensive quality score is lower than the preset threshold includes: calculating four types of indicators, including pseudo-label quantity stability, average fusion confidence change, high discrete cluster ratio, and consistency of continuous frame category and location. Calculate the overall quality score according to the preset weights; When the overall quality score reaches the threshold, it will be updated in this round and written into the stable parameters. When the overall quality score is lower than the threshold, depending on the anomaly type, skip the update, reduce the learning rate, tighten the pseudo-label threshold, or fall back to the previous stable parameter. The exception type, context parameters, pseudo-label size, and stability parameters are recorded as feedback samples.

10. A transmission line hazard target detection test adaptive system, characterized by, include: The initial inference module is configured to acquire inspection images, perform forward inference on the inspection images using a pre-trained detector, and generate an original prediction set, which includes prediction box coordinates, category labels, and confidence scores. The geometrically perceptive prior filtering module is configured to perform geometric verification on the original prediction set to obtain a geometrically reasonable prediction set; The area consistency constraint module is configured to perform spatial neighborhood clustering on the geometrically reasonable prediction set to obtain multiple candidate clusters, calculate the scale dispersion of the predicted box area within each candidate cluster, and retain only the predictions with confidence scores not lower than a set threshold for high dispersion clusters with scale dispersion exceeding a dispersion threshold to obtain a scale-consistent prediction set. The classification-aware weighted box fusion module is configured to group the scale-consistent prediction set by category label, dynamically cluster the prediction boxes within the group, perform a weighted average of the prediction box coordinates with confidence scores as weights, calculate the fusion confidence, and obtain a refined pseudo-label set. The online parameter update module is configured to use the refined pseudo-label set as a supervision signal to calculate the detection loss, and to iteratively update the model parameters of the pre-trained detector through an optimization algorithm until the number of iterations is met. During the update process, a comprehensive quality score is calculated based on the quality index of the refined pseudo-label set. When the comprehensive quality score is lower than a preset threshold, the current round of parameter update is skipped. The final inference and early warning module is configured to use the updated pre-trained detector to perform final inference on the inspection image, obtain the hidden danger detection result, and issue an early warning.

Images