Power transmission line outer skin damage intelligent identification method and system based on deep learning

Abstract

This invention relates to the field of power grid operation safety technology, and discloses a method and system for intelligent identification of transmission line outer sheath damage based on deep learning. The method includes the following steps: acquiring original images and BeiDou coordinates, completing timestamp alignment and basic preprocessing to obtain an enhanced image as the inference input; using a unified model to perform multi-task inference on the enhanced image, outputting a multi-instance inference set, and performing temperature calibration on the classification confidence, calculating the centroid / box center, normalizing the severity, and measuring uncertainty; the multi-instance inference set includes bounding boxes, segmentation masks, severity scores, classification confidence, and uncertainty. This invention can suppress the impact of positioning errors on identification and early warning in scenarios with obstruction and multipath conditions such as valleys, urban canyons, and metal structure reflections, reducing the risks of missed detections, false detections, and delayed handling, improving stability and robustness in complex weather and long corridor scenarios, and achieving all-weather, low-latency, and traceable on-site management.

Description

Technical Field

[0001] This invention relates to the field of power grid operation safety, and more specifically, to a method and system for intelligent identification of transmission line sheath damage based on deep learning. Background Technology

[0002] Damage to the outer sheath of transmission lines (including cracks, peeling, fissures, bulges, scratches, etc. on conductor anti-corrosion sheaths, insulator skirts, hardware rubber covers, and various sheaths) is one of the high-frequency hidden dangers affecting the safe and economical operation of the power grid.

[0003] Traditional methods relying on manual inspections and experience-based interpretation are prone to missed detections, false detections, and delayed responses in long corridors, multiple scenarios, and complex weather conditions. In particular, in scenarios such as valleys, urban canyons, and metal structure reflections, signal obstruction and multipath propagation amplify positioning errors, affecting the calculation of electronic fence distances and early warning responses, making it difficult to meet the on-site management requirements of "all-weather, low latency, and traceability". Summary of the Invention

[0004] This invention provides a method and system for intelligent identification of damage to the outer sheath of power transmission lines based on deep learning. It solves the technical problem in related technologies where signal obstruction and multipathing amplify positioning errors in scenarios such as valleys, urban canyons, and metal structure reflections, affecting the calculation of electronic fence distances and early warning response.

[0005] This invention provides a deep learning-based intelligent identification method for transmission line sheath damage, comprising the following steps:

[0006] S100, Data Access and Preprocessing: Acquire raw images and BeiDou coordinates, complete timestamp alignment and basic preprocessing, and obtain the enhanced image for inference input;

[0007] S200, Unified Model Inference: The unified model performs multi-task inference on the enhanced image, outputs a multi-instance inference set, performs temperature calibration on the classification confidence, calculates the centroid / box center, normalizes the severity and measures the uncertainty;

[0008] The multi-instance inference set includes bounding boxes, segmentation masks, severity scores, classification confidence, and uncertainty;

[0009] S300, Robust Positioning Filter: Based on local... The coordinates represent the state, integrating BeiDou and IMU / odometer to suppress obstruction and multipath errors;

[0010] S400, Distance Probability and Gating: Linearizes the distance from a point to a fence, calculates the mean / variance of the distance, and uses probability gating to trigger the signal.

[0011] S500, reliability weight and risk fusion: calculate reliability weight based on positioning covariance trace; back up visual estimation in case of low-quality positioning; update risk score;

[0012] S600, Dynamic Early Warning and Quality Gating: Dynamically widens the early warning distance based on DOP and covariance, and sets low-quality positioning gating signs;

[0013] S700, Historical Trajectory Adsorption and Anomaly Suppression: Position adsorption is performed based on compliant trajectories to suppress abnormal jump points and improve trajectory stability;

[0014] S800, Early Warning and Closed Loop: Uses adjusted early warning distance, probability gating and updated risk to trigger in a graded manner, and records closed loop information.

[0015] Furthermore, the data access and preprocessing steps are as follows:

[0016] S110, Time and Data Alignment: Unify the image acquisition time and positioning time to the same time base, and use linear interpolation to align missing frames;

[0017] S120, Image preprocessing and enhancement: Perform geometric correction and scene-based enhancement to obtain the enhanced image;

[0018] Enhancement operators include noise reduction, glare reduction, and gamma correction.

[0019] Furthermore, the steps of unified model inference are as follows:

[0020] S210, Model Inference Output Set: Inference is performed on the input augmented image to obtain multi-instance results;

[0021] S220, Classification Probability and Temperature Calibration: Temperature calibration is performed on the logit to obtain stable class probabilities and instance confidence;

[0022] S230, Instance segmentation and centroid localization: Calculate the centroid and bounding box center of the image plane based on the mask, which is used for subsequent visual geometric backtracking;

[0023] S240, Severity score normalization: The original severity output is normalized by scale and shift to obtain a score in the [0,1] interval;

[0024] S250, Uncertainty Measure: Model uncertainty is measured using predictive entropy or MC Dropout divergence.

[0025] Furthermore, the robust localization filter takes the following steps:

[0026] S310, State prediction: One-step prediction based on the motion model;

[0027] S320, State Update: Fusion of observations yields posterior estimates and covariance.

[0028] Furthermore, the state prediction formula is as follows:

[0029] ;

[0030] in, This refers to the state at the previous moment; For predicting the state; The covariance of the previous time step. To predict covariance; Here is the state transition matrix. This is the transpose of the state transition matrix; For process noise covariance;

[0031] The formula for state update is as follows:

[0032] ;

[0033] ;

[0034] in, For the observation matrix, This is the transpose of the observation matrix; To observe the noise covariance; For observation vectors; Kalman gain; These are the updated posterior state and covariance, respectively. For predicting the state; It is an identity matrix.

[0035] Furthermore, the steps for distance probability and gating are as follows:

[0036] S410, Distance Statistics Estimation: Estimating the mean and variance of distances based on linearization and posterior covariance;

[0037] ;

[0038] in, This is a function for finding the closest distance from a point to a fence. For posterior state estimation; These are the mean and variance of the distance, respectively. For gradient;

[0039] S420, probability-gated triggering: Based on the normal approximation, calculate the probability of entering the warning zone and compare it with the threshold;

[0040] ;

[0041] in, This is the warning distance threshold; Standard normal ; As a probability threshold, The mean distance; The distance is the standard deviation. This is a probability-triggered indicator. For indicator functions, where express .

[0042] Furthermore, the steps for integrating reliability weights and risk are as follows:

[0043] S510, Positioning Reliability Weights: Weights are constructed using the covariance trace; the larger the covariance, the smaller the weight.

[0044] ;

[0045] in, For covariance trace; For normalization; Assigning a reliability weight to the location;

[0046] S520, Visual Estimation Backoff: Based on the homography matrix, image points are mapped to ground coordinates to estimate the distance to the fence;

[0047] ;

[0048] in, This is the matrix representing the correspondence between the image and the ground. For image points; Ground coordinates; Used as a reference point for the fence; For visual distance estimation, Let L be the L2 norm, where for ;

[0049] S530, Risk Score Fusion: Weighted fusion of original risk and visual risk based on location reliability weight;

[0050] ;

[0051] in, The original comprehensive risk score; Severity; It is Sigmoid; For weighting; It is the distance scale constant; For the updated risk score, Assigning a reliability weight to the location; This is the warning distance threshold.

[0052] Furthermore, the steps for dynamic early warning and quality gating are as follows:

[0053] S610, Dynamic adjustment of warning distance: Automatically adjusts the warning distance using quality indicators;

[0054] S620, Low-quality positioning gating: Sets low-quality positioning flags based on DOP and covariance trace.

[0055] Furthermore, the steps of historical trajectory adsorption and anomaly suppression are as follows:

[0056] S710, Adsorption position calculation: Calculate the nearest point from the current state to the compliant trajectory as the adsorption position;

[0057] ;

[0058] in, For historical compliance trajectory, multiple segments; This is the current posterior position; This is the adsorption site. Furthermore... The operator that takes the minimum value; It is a norm 2;

[0059] S720, Adsorption enable gating: Adsorption is enabled when the deviation exceeds a threshold;

[0060] ;

[0061] in, The adsorption threshold; Indicates adsorption is enabled.

[0062] This invention also proposes a deep learning-based intelligent identification system for transmission line sheath damage, comprising:

[0063] Module 1: This module is responsible for continuously collecting image and coordinate data from high-definition visualization devices and Beidou positioning terminals, completing timestamp unification and frame-level alignment, and performing geometric correction and scene enhancement;

[0064] Module 2: This module runs a multi-task unified model on the algorithm host to detect and segment targets related to skin damage in the input image, and outputs the severity score, recognition confidence, and model uncertainty for each instance.

[0065] Module 3: This module uses a local coordinate system as its state, integrates BeiDou positioning with auxiliary sensor information such as IMU / odometer, and adopts a prediction-update filtering process to suppress positioning fluctuations caused by occlusion and multipath.

[0066] Module 4: This module estimates the distance distribution from the target to the fence based on the posterior of the location and the geometry of the electronic fence, and performs probabilistic gating based on the probability of entering the warning area;

[0067] Module 5: This module generates positioning reliability weights based on the quality index of positioning covariance, and reverts to visual geometric distance estimation under low-quality positioning conditions; then, the original risk score and visual risk are weighted and fused to obtain the updated comprehensive risk.

[0068] Module Six: This module combines quality indicators such as DOP and covariance to dynamically adjust the warning distance of the electronic fence and set low-quality positioning gate markers to participate in subsequent triggering logic;

[0069] Module 7: This module uses recent compliant trajectories to snap to the current location, suppressing abnormal jumps and drift phenomena;

[0070] Module 8: This module integrates the adjusted warning distance, probability gating results, low-quality location markers, and updated comprehensive risks to execute tiered triggering and message distribution.

[0071] The beneficial effects of this invention are as follows:

[0072] This invention constructs robust positioning by unifying inference and temperature calibration, uncertainty measurement, and using extended Kalman filtering with BeiDou / IMU fusion. It also uses visual geometric backoff, positioning reliability weighting, and probability gating to collaboratively correct the distance and triggering conditions of the electronic fence. At the same time, it introduces dynamic adjustment of warning distance based on DOP and historical trajectory adsorption to stabilize the trajectory, forming a hierarchical warning and closed-loop recording. In scenarios with obstruction and multipath, such as valleys, urban canyons, and metal structure reflections, it effectively suppresses the impact of positioning errors on identification and warning, reduces the risk of missed detections, false detections, and delayed handling, and improves the stability and robustness in complex weather and long corridor scenarios, achieving all-weather, low-latency, and traceable on-site management. Attached Figure Description

[0073] Figure 1 This is a flowchart of the intelligent identification method for transmission line outer sheath damage based on deep learning according to the present invention;

[0074] Figure 2 This is a structural block diagram of the intelligent identification system for transmission line sheath damage based on deep learning according to the present invention;

[0075] Figure 3 This is a schematic diagram of data and process in an example of the present invention;

[0076] Figure 4 This is a schematic diagram of the electronic fence distance in an example of the present invention;

[0077] Figure 5This is the probability gating curve in an example of the present invention;

[0078] Figure 6 This is a visual backtracking mapping diagram in an example of the present invention;

[0079] Figure 7 This is a schematic diagram of trajectory adsorption in an example of the present invention. Detailed Implementation

[0080] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0081] like Figures 1-7 As shown, a deep learning-based intelligent identification method for transmission line outer sheath damage includes...

[0082] S100, Data Access and Preprocessing: Image Acquisition with BeiDou coordinates After completing timestamp alignment and basic preprocessing, the inference input is obtained. ;

[0083] In one embodiment of the present invention, the following steps are specifically included:

[0084] S110, Time and Data Alignment: Unify the image acquisition time and positioning time to the same time base, and use linear interpolation to align missing frames.

[0085] formula:

[0086] ;

[0087] in, Image frame time; To correct the time for positioning; For time offset; Original location; To approximate the positioning speed, To correct the positioning.

[0088] S120, Image Preprocessing and Enhancement: Performing geometric correction and scene-based enhancement (noise reduction, glare removal, gamma correction) to obtain... .

[0089] formula:

[0090] ;

[0091] in, Original image; For the enhanced image; To enhance the set of operators (denoising / brightening / distortion correction).

[0092] S200, Unified Model Inference (Detection-Segmentation-Evaluation): The unified model performs multi-task inference on the enhanced image, outputs a multi-instance inference set, performs temperature calibration on the classification confidence, calculates the centroid / box center, normalizes the severity, and measures uncertainty.

[0093] In one embodiment of the present invention, the following steps are specifically included:

[0094] S210, Model Inference Output Set: For input... Reasoning is performed to obtain multiple instance results.

[0095] formula:

[0096] ;

[0097] in, A unified model for multiple tasks; These are model parameters; The number of instances; For bounding boxes; For segmentation mask; Severity rating (0-1); For classification confidence; It is uncertain.

[0098] S220, Classification Probability and Temperature Calibration: Temperature calibration is performed on the logit to obtain stable class probabilities and instance confidence.

[0099] formula:

[0100] ;

[0101] ;

[0102] ;

[0103] in, For the i-th instance, the logit for the c-th class, For the i-th instance, the logit for all categories; Temperature calibration coefficient ( ); The calibrated class probability; For the instance with the highest class confidence, To find the summation variable, iterate through all c categories. This is an operation to take the maximum value of the category dimension. The minimum confidence threshold; For filtering indication, This is an indicator function.

[0104] S230, Instance Segmentation and Centroid Localization: Calculate the centroid and bounding box center of the image plane based on the mask, which is used for subsequent visual geometric backtracking.

[0105] formula:

[0106] ;

[0107] ;

[0108] in, For the i-th instance in pixel Mask value at the location; These are pixel coordinates; For masking the centroid; Center of the bounding box; The bounding box coordinates are, i.e. These are the upper and lower bounds of the horizontal bounding box. These are the upper and lower limits of the vertical bounding box.

[0109] S240, Severity Score Normalization: The original severity output is normalized by scale and shift to obtain a score in the [0,1] interval.

[0110] formula:

[0111] ;

[0112] in, This is the original severity output; These are translation parameters; For scale parameters ( ); For the Sigmoid function; To normalize the severity output, .

[0113] S250, Uncertainty Measure: The model uncertainty is measured using predictive entropy or divergence from MC Dropout (Monte Carlo Dropout).

[0114] Formula (prediction entropy):

[0115] ;

[0116] Formula (MC Dropout):

[0117] ;

[0118] in, The number of MC samples; Let be the class probability of the m-th sample; The probability mean; The Kullback–Leibler divergence; This is a measure of uncertainty (the larger the value, the greater the uncertainty). Another... The category probability after temperature calibration; This refers to uncertainty based on prediction entropy.

[0119] S300, Robust Localization Filter (EKF, Extended Kalman Filter): Based on local... The coordinates represent the state, integrating BeiDou and IMU / odometer to suppress occlusion and multipath errors.

[0120] In one embodiment of the present invention, the following steps are specifically included:

[0121] S310, State prediction: One-step prediction based on the motion model.

[0122] formula:

[0123] ;

[0124] in, This refers to the state at the previous moment; For predicting the state; The covariance of the previous time step. To predict covariance; This is the state transition matrix; Let be the process noise covariance.

[0125] S320, State Update: Fusion of observations yields posterior estimates and covariance.

[0126] formula:

[0127] ;

[0128] ;

[0129] in, The observation matrix; To observe noise covariance (including multipath effects); For observation vectors; Kalman gain; The updated posterior state and covariance, For predicting the state; To predict covariance; It is an identity matrix.

[0130] S400, Distance Probability and Gating: Linearizes the distance from a point to a fence, calculates the mean / variance of the distance, and uses probability gating for triggering.

[0131] In one embodiment of the present invention, the following steps are specifically included:

[0132] S410, Distance Statistics Estimation: Estimating the mean and variance of distances based on linearization and posterior covariance.

[0133] formula:

[0134] ;

[0135] in, This is a function for finding the closest distance from a point to a fence. For posterior state estimation; These are the mean and variance of the distance, respectively. For gradient.

[0136] S420, probability-gated trigger: Based on the normal approximation, calculate the probability of entering the warning zone and compare it with the threshold.

[0137] formula:

[0138] ;

[0139] in, This is the warning distance threshold; Standard normal (Cumulative Distribution Function) A probability threshold (e.g., 0.9). The mean distance; The distance is the standard deviation. This is a probability-triggered indicator. This is an indicator function.

[0140] S500, reliability weight and risk fusion: calculate reliability weight based on positioning covariance trace; back up visual estimation in case of low-quality positioning; update risk score.

[0141] In one embodiment of the present invention, the following steps are specifically included:

[0142] S510, Positioning Reliability Weights: Weights are constructed using covariance traces; the larger the covariance, the smaller the weight.

[0143] formula:

[0144] ;

[0145] in, For covariance trace; For normalization; To determine the reliability weight.

[0146] S520, Visual Estimation Backoff: Based on the homography matrix, image points are mapped to ground coordinates to estimate the distance to the fence.

[0147] formula:

[0148] ;

[0149] in, This is the homography matrix from the image to the ground; For image points; Ground coordinates; Used as a reference point for the fence; For visual distance estimation, Let L be the L2 norm, where for .

[0150] S530, Risk Score Fusion: Weighted fusion of original risk and visual risk based on location reliability weight.

[0151] formula:

[0152] ;

[0153] in, The original comprehensive risk score; Severity; For Sigmoid, where for ; For weighting; It is the distance scale constant; For the updated risk score, Assigning a reliability weight to the location; This is the warning distance threshold.

[0154] S600, Dynamic Early Warning and Quality Gating: Based on DOP (Dilution of Precision) and covariance, the early warning distance is dynamically widened, and a low-quality positioning gating flag is set.

[0155] In one embodiment of the present invention, the following steps are specifically included:

[0156] S610, Dynamic adjustment of warning distance: The warning distance is automatically adjusted using quality indicators.

[0157] formula:

[0158] ;

[0159] in, This is the adjusted warning distance; Amplify the base volume; For the current DOP; For reference DOP, This is the original warning distance threshold; This serves as a normalization metric for positioning reliability.

[0160] S620, Low-quality positioning gating: Sets low-quality positioning flags based on DOP and covariance trace.

[0161] formula:

[0162] ;

[0163] in, DOP threshold; The covariance trace threshold; This is a low-quality positioning indicator. This is the current positioning geometric precision factor; For covariance trace.

[0164] S700, Historical Trajectory Adsorption and Anomaly Suppression: Position adsorption is performed based on compliant trajectories to suppress abnormal jumps and improve trajectory stability.

[0165] In one embodiment of the present invention, the following steps are specifically included:

[0166] S710, Adsorption position calculation: Calculate the nearest point from the current state to the compliant trajectory as the adsorption position.

[0167] formula:

[0168] ;

[0169] in, For historical compliance trajectory, multiple segments; This is the current posterior position; This is the adsorption site. Furthermore... The operator that takes the minimum value; It is a 2-norm.

[0170] S720, Adsorption Enable Gating: Adsorption is enabled when the deviation exceeds a threshold.

[0171] ;

[0172] in, The adsorption threshold; Indicates adsorption is enabled.

[0173] S800, Early Warning and Closed Loop: Uses adjusted early warning distance, probability gating and updated risk to trigger in a graded manner, and records closed loop information.

[0174] In one embodiment of the present invention, the following steps are specifically included:

[0175] S810, Hierarchy and Triggering: Comprehensive Consideration , , and Trigger an early warning.

[0176] formula:

[0177] ;

[0178] in, This is the final trigger indication; The risk threshold is set to the general level (configurable). This is a probability-triggered indicator; This is the updated risk score.

[0179] S820, closed-loop recording: writes quality gate control flags, adsorption status, risk scores and visual backtracking logs into the system for easy review and auditing.

[0180] Based on the aforementioned deep learning-based intelligent identification method for transmission line sheath damage, and to execute the steps in the method, the following deep learning-based intelligent identification system for transmission line sheath damage is also proposed. The system includes the following modules:

[0181] Module 1, Data Access and Preprocessing: This module is responsible for continuously acquiring image and coordinate data from high-definition visualization devices and BeiDou positioning terminals, completing timestamp unification and frame-level alignment, and performing geometric correction and scene enhancement (such as noise reduction, glare removal, gamma and color correction). The module outputs standardized image inputs and synchronized positioning information that can be used for downstream inference, ensuring good usability and consistency of data even in complex environments.

[0182] Module 2, Unified Model Inference (Detection-Segmentation-Evaluation): This module runs a multi-task unified model on the algorithm host, detecting and segmenting targets with damaged outer skin in the input image. It outputs a severity score, recognition confidence, and model uncertainty for each instance. The module results are used for subsequent risk fusion and alarm gating, supporting the entire chain from visual recognition to business processing.

[0183] Module 3, Robust Positioning Filter (EKF): This module uses a local coordinate system as its state, integrating BeiDou positioning with auxiliary sensor information such as IMU / odometer, and employs a prediction-update filtering process to suppress positioning fluctuations caused by occlusion and multipath propagation. The module continuously outputs quality indicators such as posterior position and covariance, providing a reliable positioning foundation for subsequent electronic fence distance calculation and quality gating.

[0184] Module 4, Distance Probability and Gating: This module estimates the distance distribution from the target to the fence based on the posterior of the positioning and the geometry of the electronic fence, and performs probabilistic gating based on the probability of entering the warning area. Compared to a fixed threshold, this module can maintain a robust triggering strategy in situations where positioning uncertainty increases, reducing false alarms and false negatives.

[0185] Module 5, Reliability Weighting and Risk Fusion: This module generates positioning reliability weights based on the quality index of positioning covariance and reverts to visual geometric distance estimation under low-quality positioning conditions. Subsequently, the original risk score and visual risk are weighted and fused to obtain an updated comprehensive risk. This module ensures that risk assessment remains stable and reliable despite fluctuations in positioning quality, guaranteeing timely identification of high-risk events.

[0186] Module Six, Dynamic Early Warning and Quality Gating: This module combines quality indicators such as DOP and covariance to dynamically adjust the warning distance of the electronic fence and set low-quality positioning gating flags to participate in subsequent triggering logic. The module automatically increases the safety margin in complex scenarios such as communication obstruction and urban canyons, reducing false triggering caused by special environments.

[0187] Module 7, Historical Trajectory Absorption and Anomaly Suppression: This module uses recent compliant trajectories to absorb the current location, suppressing abnormal jumps and drift phenomena. When the deviation exceeds a set threshold, the absorption strategy is automatically activated, and the relevant status is written to the log, improving the continuity of the trajectory and the business side's trust in the location quality.

[0188] Module 8, Early Warning and Closed Loop: This module integrates the adjusted early warning distance, probability gating results, low-quality positioning markers, and updated comprehensive risks to execute tiered triggering and message distribution; at the same time, it records quality and gating labels, visual rollback and adsorption logs on the WEB / APP terminal to achieve verifiable and auditable closed-loop management, supporting rapid on-site personnel handling and post-event tracking and analysis.

[0189] Based on the steps and modules of the above-mentioned deep learning-based intelligent identification method for transmission line sheath damage, the following example is applied, with the scenario overview of the example as follows:

[0190] In an urban canyon environment, construction machinery is operating near a power tower; visual monitoring is normal, but BeiDou positioning fluctuates significantly due to obstruction and multipath effects. The system achieves a complete process from identification to warning and then to closed-loop control through robust positioning and probabilistic gating, visual backoff, dynamic early warning and trajectory snapping.

[0191] In this example, the application flow is as follows:

[0192] like Figure 3 As shown in the data and process diagram, the data access and preprocessing steps are as follows: high-resolution images and BeiDou coordinates are acquired, and timestamp unification and frame-level alignment are completed. Geometric correction and scene enhancement are performed on the images, and standardized images and synchronous positioning information are output for inference.

[0193] Unified Model Inference: Detection-Segmentation-Evaluation: A multi-task unified model is run on the algorithm host to identify targets such as construction machinery, and output instance segmentation masks and severity scores, confidence and uncertainty for subsequent risk fusion.

[0194] Robust positioning filtering: Integrates BeiDou and IMU / odometer for prediction and update, outputs posterior position and covariance and other quality indicators, and suppresses positioning jumps caused by occlusion and multipath.

[0195] Probabilistic Distance and Probability Gating: Based on Location Posterior and Geometric Fence, such as Figure 4 The diagram illustrates the distance distribution to the electronic fence, estimating the distance distribution to the fence and implementing gating based on the probability of entering the warning zone. Figure 5 As shown in the probability gating curve, Φ(Z) in the figure represents the standard normal CDF, where Z is the normalized distance. * This is a gated distance to avoid misjudgments under high uncertainty environments with a fixed threshold.

[0196] Reliability weights are integrated with risk: reliability weights are generated based on the positioning covariance, such as... Figure 6 As shown in the visual backtracking mapping diagram, image point x is mapped to ground point X through the homography matrix, and the nearest distance to the fence is estimated. When the localization is of poor quality, it backtracks to the visual geometric distance estimation. The original risk and visual risk are weighted and fused to obtain a more robust comprehensive risk.

[0197] Dynamic early warning and quality gating: The early warning distance is dynamically widened based on DP and covariance, and a low-quality positioning gating flag is set to participate in the triggering logic to improve the safety margin in complex scenarios.

[0198] Trajectory adsorption and anomaly suppression: such as Figure 7 As shown in the trajectory adsorption diagram, the current position is projected onto the nearest point of the historical trajectory to suppress abnormal jumps. Referring to recent compliant trajectories, the current location is adsorbed to suppress abnormal jumps, and the adsorption status is recorded to support auditing and review.

[0199] Early warning and closed loop: The comprehensive risk of the adjusted early warning distance, probability gating, quality gating and update is classified and triggered; the early warning is issued to the APP / WEB terminal and written to the quality and gating, visual rollback and adsorption logs to form an auditable closed loop.

[0200] In urban canyon environments with significant BeiDou obstruction and multipath propagation, the system, with robust positioning and probabilistic gating as its core, supplemented by visual backtracking, dynamic early warning, and trajectory snapping strategies, achieves reliable early warning and an auditable closed loop for construction machinery approaching electronic fences. In complex scenarios, the collaborative design of "quality perception, policy adaptation, and multi-source fusion" enhances recognition reliability and on-site usability.

[0201] The embodiments of the present invention have been described above, but the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention, all of which are within the protection scope of the present invention.

Claims

1. A deep learning-based intelligent identification method for transmission line sheath damage, characterized in that, Includes the following steps: S100, Data Access and Preprocessing: Acquire raw images and BeiDou coordinates, complete timestamp alignment and basic preprocessing, and obtain the enhanced image for inference input; S200, Unified Model Inference: The unified model performs multi-task inference on the enhanced image, outputs a multi-instance inference set, performs temperature calibration on the classification confidence, calculates the centroid / box center, normalizes the severity and measures the uncertainty; The multi-instance inference set includes bounding boxes, segmentation masks, severity scores, classification confidence, and uncertainty; S300, Robust Positioning Filter: Based on local... The coordinates represent the state, integrating BeiDou and IMU / odometer to suppress obstruction and multipath errors; S400, Distance Probability and Gating: Linearizes the distance from a point to a fence, calculates the mean / variance of the distance, and uses probability gating to trigger the signal. S500, reliability weight and risk fusion: calculate reliability weight based on positioning covariance trace; back up visual estimation in case of low-quality positioning; update risk score; S600, Dynamic Early Warning and Quality Gating: Dynamically widens the early warning distance based on DOP and covariance, and sets low-quality positioning gating signs; S700, Historical Trajectory Adsorption and Anomaly Suppression: Position adsorption is performed based on compliant trajectories to suppress abnormal jump points and improve trajectory stability; S800, Early Warning and Closed Loop: Uses adjusted early warning distance, probability gating and updated risk to trigger in a graded manner, and records closed loop information.

2. The intelligent identification method for transmission line sheath damage based on deep learning according to claim 1, characterized in that, The steps for data access and preprocessing are as follows: S110, Time and Data Alignment: Unify the image acquisition time and positioning time to the same time base, and use linear interpolation to align missing frames; S120, Image preprocessing and enhancement: Perform geometric correction and scene-based enhancement to obtain the enhanced image; Enhancement operators include noise reduction, glare reduction, and gamma correction.

3. The intelligent identification method for transmission line sheath damage based on deep learning according to claim 1, characterized in that, The steps of unified model inference are as follows: S210, Model Inference Output Set: Inference is performed on the input augmented image to obtain multi-instance results; S220, Classification Probability and Temperature Calibration: Temperature calibration is performed on the logit to obtain stable class probabilities and instance confidence; S230, Instance segmentation and centroid localization: Calculate the centroid and bounding box center of the image plane based on the mask, which is used for subsequent visual geometric backtracking; S240, Severity score normalization: The original severity output is normalized by scale and shift to obtain a score in the [0,1] interval; S250, Uncertainty Measure: Model uncertainty is measured using predictive entropy or MC Dropout divergence.

4. The intelligent identification method for transmission line sheath damage based on deep learning according to claim 1, characterized in that, The steps of robust localization filtering are as follows: S310, State prediction: One-step prediction based on the motion model; S320, State Update: Fusion of observations yields posterior estimates and covariance.

5. The intelligent identification method for transmission line sheath damage based on deep learning according to claim 4, characterized in that, The state prediction formula is as follows: ； in, This refers to the state at the previous moment; For predicting the state; The covariance of the previous time step. To predict covariance; Here is the state transition matrix. This is the transpose of the state transition matrix; For process noise covariance; The formula for state update is as follows: ； ； in, For the observation matrix, This is the transpose of the observation matrix; To observe the noise covariance; For observation vectors; Kalman gain; These are the updated posterior state and covariance, respectively. For predicting the state; It is an identity matrix.

6. The intelligent identification method for transmission line sheath damage based on deep learning according to claim 1, characterized in that, The steps for distance probability and gating are as follows: S410, Distance Statistics Estimation: Estimating the mean and variance of distances based on linearization and posterior covariance; ； in, This is a function for finding the closest distance from a point to a fence. For posterior state estimation; These are the mean and variance of the distance, respectively. For gradient; S420, probability-gated triggering: Based on the normal approximation, calculate the probability of entering the warning zone and compare it with the threshold; ； in, This is the warning distance threshold; Standard normal ; As a probability threshold, The mean distance; The distance is the standard deviation. This is a probability-triggered indicator. For indicator functions, where express .

7. The intelligent identification method for transmission line sheath damage based on deep learning according to claim 1, characterized in that, The steps for integrating reliability weights and risk are as follows: S510, Positioning Reliability Weights: Weights are constructed using the covariance trace; the larger the covariance, the smaller the weight. ； in, For covariance trace; For normalization; Assigning a reliability weight to the location; S520, Visual Estimation Backoff: Based on the homography matrix, image points are mapped to ground coordinates to estimate the distance to the fence; ； in, This is the homography matrix from the image to the ground. The homography matrix is ​​a 3×3 matrix that describes the projection transformation relationship between two planes. For image points; Ground coordinates; Used as a reference point for the fence; For visual distance estimation, Let L be the L2 norm, where for ; S530, Risk Score Fusion: Weighted fusion of original risk and visual risk based on location reliability weight; ； in, The original comprehensive risk score; Severity; It is Sigmoid; For weighting; It is the distance scale constant; For the updated risk score, Assigning a reliability weight to the location; This is the warning distance threshold.

8. The intelligent identification method for transmission line sheath damage based on deep learning according to claim 1, characterized in that, The steps for dynamic early warning and quality gating are as follows: S610, Dynamic adjustment of warning distance: Automatically adjusts the warning distance using quality indicators; S620, Low-quality positioning gating: Sets low-quality positioning flags based on DOP and covariance trace.

9. The intelligent identification method for transmission line sheath damage based on deep learning according to claim 1, characterized in that, The steps of historical trajectory adsorption and anomaly suppression are as follows: S710, Adsorption position calculation: Calculate the nearest point from the current state to the compliant trajectory as the adsorption position; ； in, For historical compliance trajectory, multiple segments; This is the current posterior position; For adsorption sites, in addition, The operator that takes the minimum value; It is a norm 2; S720, Adsorption enable gating: Adsorption is enabled when the deviation exceeds a threshold; ； in, The adsorption threshold; Indicates adsorption is enabled.

10. A deep learning-based intelligent identification system for damage to the outer sheath of power transmission lines, characterized in that, The method for performing the steps in the deep learning-based intelligent identification method for transmission line sheath damage as described in any one of claims 1-9 includes: Module 1: This module is responsible for continuously collecting image and coordinate data from high-definition visualization devices and Beidou positioning terminals, completing timestamp unification and frame-level alignment, and performing geometric correction and scene enhancement; Module 2: This module runs a multi-task unified model on the algorithm host to detect and segment targets related to skin damage in the input image, and outputs the severity score, recognition confidence, and model uncertainty for each instance. Module 3: This module uses a local coordinate system as its state, integrates BeiDou positioning with auxiliary sensor information such as IMU / odometer, and adopts a prediction-update filtering process to suppress positioning fluctuations caused by occlusion and multipath. Module 4: This module estimates the distance distribution from the target to the fence based on the posterior of the location and the geometry of the electronic fence, and performs probabilistic gating based on the probability of entering the warning area; Module 5: This module generates positioning reliability weights based on the quality index of positioning covariance, and reverts to visual geometric distance estimation under low-quality positioning conditions; then, the original risk score and visual risk are weighted and fused to obtain the updated comprehensive risk. Module Six: This module combines quality indicators such as DOP and covariance to dynamically adjust the warning distance of the electronic fence and set low-quality positioning gate markers to participate in subsequent triggering logic; Module 7: This module uses recent compliant trajectories to snap to the current location, suppressing abnormal jumps and drift phenomena; Module 8: This module integrates the adjusted warning distance, probability gating results, low-quality location markers, and updated comprehensive risks to execute tiered triggering and message distribution.

Images