Power equipment defect detection system and method based on deep learning

Abstract

The invention relates to the technical field of electrical equipment defect detection, in particular to an electrical equipment defect detection system and method based on deep learning, which are characterized in that a three-dimensional model of a power grid region is constructed, and based on historical defect data, a neural network model is adopted to mark an importance score of an inspection object in the three-dimensional model; quantitative evaluation of the equipment fault risk level is completed, and the matching degree of inspection resources and defect risk distribution is improved. A power grid three-dimensional model and importance scores are combined, a reinforcement learning model is utilized to construct an unmanned aerial vehicle inspection route planning strategy, the inspection route comprehensively considers power equipment defect risks and space factors in the planning stage, task allocation is optimized, and the problem that a static route cannot adapt to equipment changes is solved. In the inspection execution process, the unmanned aerial vehicle is dispatched according to the planning strategy, and the image flow is synchronously acquired for defect detection, so that the linkage of the inspection action and the detection process is realized, and the response efficiency and the detection quality of potential defects are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment defect detection, and in particular to a power equipment defect detection system and method based on deep learning. Background Art

[0002] Existing drone inspections often employ a remote patrol model, conducting continuous patrols along pre-defined power lines, substation boundaries, or fixed routes. This model lacks targeted risk assessment and prioritization of inspection targets. This remote patrol model typically plans routes based on geographic connectivity or the shortest path, ignoring the significant differences in operating status, historical defect probability, and safety importance of various power equipment. This inadequate consideration of the importance of inspection targets leads to a mismatch between inspection resource investment and fault detection benefits, resulting in wasted flight resources and reduced detection effectiveness.

[0003] Secondly, existing drone inspection routes are generally based on static route planning, and the flight route and task points are usually determined once before the inspection mission begins. They do not have the ability to dynamically optimize the inspection route according to the real-time detection status of the power equipment during flight. Most current inspection systems only support image acquisition or observation tasks at preset waypoints and lack linkage with real-time defect detection systems. When a drone discovers new potential defect targets during the inspection process, it is unable to adjust the remaining inspection plan based on real-time feedback, such as increasing close-range flights or hovering observations of suspected defective areas, which makes it easy for missed inspections or detection failures. In addition, because the inspection route cannot be flexibly adjusted as the equipment status changes, the drone may continue to fly on the original route and will not be able to respond effectively even when faced with new abnormal signals, affecting the quality of defect detection of power equipment in the power grid. Summary of the Invention

[0004] To solve the above problems, the present invention provides a power equipment defect detection system and method based on deep learning.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A method for detecting defects in power equipment based on deep learning, comprising:

[0007] S1. Construct a three-dimensional model of the power grid area;

[0008] S2. Based on the historical defect data of power equipment in the power grid area, a neural network model is used to mark the importance scores of inspection objects in the three-dimensional model of the power grid area;

[0009] S3. Based on the three-dimensional model of the power grid area, a UAV inspection route planning strategy is constructed through a reinforcement learning model;

[0010] S4. Schedule the drone for inspection according to the drone inspection route planning strategy, and simultaneously obtain the inspection image stream to detect defects in power equipment.

[0011] Furthermore, the S1 includes the following steps:

[0012] Use several drones equipped with lidar and cameras to autonomously scan the power grid area, obtaining laser point cloud data and two-dimensional image sequences of the power grid area;

[0013] Performing filtering and registration processing based on the laser point cloud data to obtain a preliminary three-dimensional point cloud model of the power grid area;

[0014] The two-dimensional image sequence is subjected to multi-view image fusion processing to obtain a texture feature image of the power grid area, and the texture feature image is mapped to the preliminary three-dimensional point cloud model to obtain a three-dimensional model of the power grid area.

[0015] Furthermore, the S2 includes the following steps:

[0016] Based on the historical defect data of power equipment in the power grid area, a clustering algorithm is used to perform spatial hotspot analysis to obtain the initial hotspot data of the high-incidence areas of conventional defects in the power grid area;

[0017] Extract the locations of unrepaired defects and uninspected locations based on the last inspection result data to obtain the remaining defect data of the last inspection cycle;

[0018] spatially fusing the initial hotspot data with the defect legacy data to obtain a hotspot fusion data set;

[0019] Inputting the hotspot fusion data set into a convolutional neural network for training to obtain a scoring model for marking the importance score of the inspection object;

[0020] Based on the scoring model, a prediction scoring process is performed on the inspection objects in the three-dimensional model of the power grid area to obtain a three-dimensional power grid model marked with the importance scores of the inspection objects.

[0021] Furthermore, the convolutional neural network is constructed by the following steps:

[0022] Based on the hotspot fusion dataset, the location coordinates, historical defect occurrence counts, most recent inspection defect detection results, and equipment type information of each inspection object are extracted to construct a training dataset containing inspection object spatial features, historical defect features, and type features;

[0023] Based on the training data set, the training data is divided into a training set and a validation set according to the importance labels corresponding to the inspection objects;

[0024] Inputting the training set into a convolutional neural network, performing feature extraction through a convolutional layer, and obtaining a multi-scale spatial feature map;

[0025] Performing pooling dimensionality reduction processing on the multi-scale spatial feature map to obtain a compressed feature vector;

[0026] Input the compressed feature vector into the fully connected layer, perform regression training, and output the importance score of the corresponding inspection object;

[0027] The prediction accuracy of the trained convolutional neural network is verified using the validation set.

[0028] Furthermore, the step S3 includes the following steps:

[0029] Based on the three-dimensional model of the power grid area, the spatial location and corresponding importance score of each inspection object are extracted. The inspection task state space is constructed by combining the initial position, power state, flight speed and defect confirmation status of each drone.

[0030] Based on the inspection task state space, an action space is defined as an action set for each UAV to select a target inspection object in the next step, and the action set includes action options of heading to the inspection object with different headings and different ranges;

[0031] The inspection task state space and the action space are input into the reinforcement learning model, and the next inspection target point of each drone is output through action decision-making to obtain a preliminary drone inspection path sequence;

[0032] Based on the preliminary UAV inspection path sequence, the inspection path strategy is optimized through a reward function. The reward value of the reward function is calculated based on the inspection time, the number of defect targets found, the missed detection rate, and the clarity of the defect image;

[0033] Based on the strategy optimization results, the reinforcement learning model parameters are updated and a real-time drone inspection route planning strategy is generated.

[0034] Furthermore, the reward function is as follows:

[0035]

[0036] Among them, R is the reward value; T is the cumulative time it takes for the drone to complete the inspection task; T ref is the preset reference time; D is the number of defect targets successfully found during the inspection process; L is the missed detection rate of the inspection task; Q is the clarity score of the defect image collected close to the defect target; Q max is the maximum reference value of the clarity score; α1, α2 and α4 are the reward coefficients; α3 is the penalty coefficient.

[0037] Furthermore, the S4 includes the following steps:

[0038] S41, based on the inspection route planning strategy generated in real time, extracting the inspection action sequence of each drone, and generating an inspection task instruction set including the inspection target location, action instructions and execution order;

[0039] S42. Based on the inspection task instruction set, each UAV is dispatched and controlled to make the UAV fly autonomously according to the action sequence and synchronously collect flight status data and inspection image stream data at each time step;

[0040] S43, inputting the inspection image stream data into a defect detection model to perform real-time defect detection, obtaining a defect detection result of the current time step, and updating the inspection task state space based on the defect detection result;

[0041] S44. Based on the updated inspection task state space, re-input the reinforcement learning model, generate an inspection route planning strategy, and update the inspection task instruction set; if the inspection task instruction set is not empty, execute S42; if the inspection task instruction set is empty, execute S45;

[0042] S45. According to the inspection defect detection results and the inspection target location information, the defect location is marked in the three-dimensional model of the power grid area, and the defect image, equipment number and detection time are associated and stored to generate an inspection data record.

[0043] Furthermore, the defect detection model is deployed on a cloud server that is communicatively connected to the drone.

[0044] Furthermore, the defect detection model is constructed by the following steps:

[0045] Based on the historical image data of power equipment inspections in the power grid, original image samples are extracted and annotated to obtain training and verification datasets containing defect target categories and location box annotations;

[0046] Based on the training data set, performing size normalization, color normalization and data enhancement processing on the image samples to obtain preprocessed training image data;

[0047] Build the Transformer network model and initialize the network parameters;

[0048] Inputting the preprocessed training image data into the Transformer network model, and optimizing the model parameters by jointly training with weighted cross entropy loss and bounding box regression loss;

[0049] The detection accuracy and positioning accuracy of the Transformer network model are evaluated based on the validation dataset, and the model hyperparameters are adjusted according to the evaluation results to complete the defect detection model construction.

[0050] A power equipment defect detection system based on deep learning, applied to any of the aforementioned power equipment defect detection methods based on deep learning, comprising:

[0051] The power grid model module is used to build a three-dimensional model of the power grid area;

[0052] The inspection marking module is used to mark the importance scores of inspection objects on the three-dimensional model of the power grid area through a neural network model based on the historical defect data of power equipment in the power grid area;

[0053] An intelligent planning module, which uses a reinforcement learning model to build a UAV inspection route planning strategy based on the importance scores of the inspection objects and the three-dimensional model of the power grid area;

[0054] The defect detection module is used to schedule drone inspections according to the drone inspection route planning strategy and simultaneously obtain inspection image streams for power equipment defect detection.

[0055] The beneficial effects of the present invention are as follows: the present invention realizes accurate modeling of the spatial distribution of power equipment by constructing a three-dimensional model of the power grid area, thereby avoiding the waste of resources and inspection blind spot problems caused by the existing route planning model based on line continuity. Based on the historical defect data of power equipment in the power grid area, a neural network model is used to mark the importance score of the inspection object in the three-dimensional model of the power grid area, and complete the quantitative assessment and classification of the equipment failure risk level, which solves the problem of lack of task priority sorting based on equipment operation risk in the existing inspection process, and improves the matching degree between inspection resources and defect risk distribution. Based on the three-dimensional model of the power grid area and the importance score information of the inspection object, a reinforcement learning model is used to construct a drone inspection route planning strategy, so that the inspection path can comprehensively consider factors such as equipment importance, spatial location and flight cost in the planning stage, optimize the inspection task allocation and path selection, and overcome the limitations of the existing inspection path being static and fixed and unable to adapt to changes in power equipment. During the inspection execution phase, drones are dispatched for inspections based on the inspection route planning strategy, and inspection image streams are simultaneously acquired to perform power equipment defect detection. This enables the linkage between inspection actions and the defect detection process, improving the real-time response capability to potential defects, effectively reducing the missed detection rate, and increasing the drone's attention to power equipment defect issues. This allows for higher image quality for defects and improved detection accuracy. Through the aforementioned technical means, dynamic identification of the importance of inspection objects, intelligent optimization of inspection routes, and detection linkage control of inspection actions are effectively achieved during the power equipment inspection process. This significantly improves inspection efficiency, defect detection rate, and detection quality, overcoming the shortcomings of existing drone inspection systems in task allocation, path planning, and defect detection response. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flowchart of the steps of a method for detecting defects in power equipment based on deep learning in the present invention.

[0057] Figure 2 It is a step flow chart of step S4 in the present invention. DETAILED DESCRIPTION

[0058] See also Figure 1-Figure 2 As shown, the present invention relates to a method for detecting defects in power equipment based on deep learning, comprising:

[0059] S1. Construct a three-dimensional model of the power grid area;

[0060] S2. Based on the historical defect data of power equipment in the power grid area, a neural network model is used to mark the importance scores of inspection objects in the three-dimensional model of the power grid area;

[0061] S3. Based on the three-dimensional model of the power grid area, a UAV inspection route planning strategy is constructed through a reinforcement learning model;

[0062] S4. Schedule the drone for inspection according to the drone inspection route planning strategy, and simultaneously obtain the inspection image stream to detect defects in power equipment.

[0063] It should be noted that in order to detect surface defects that may exist on the exterior surfaces of power equipment such as transformers, circuit breakers, disconnectors, transmission line towers and lightning arresters in the power grid area, including surface anomalies such as cracks, corrosion, rust, damage, insulation contamination and foreign matter attachment, this embodiment first uses multiple drones equipped with lidar and high-resolution visible light cameras to autonomously scan the power grid area, and collects refined point cloud data and surface texture images for key inspection locations such as transformer boxes, circuit breaker contacts, disconnector connecting rods, transmission line tower structures and lightning arrester casings. Through point cloud filtering, alignment and texture fusion, a high-precision three-dimensional model of the power grid is generated, and the surface features of various types of power equipment are fully restored, providing a spatial basis for subsequent defect detection and inspection route planning. Compared with traditional route patrol methods, this solution can accurately cover areas where the equipment surface structure is complex and prone to defects such as cracks and corrosion, significantly improving inspection coverage and modeling accuracy. Based on the completed three-dimensional power grid model and combined with historical maintenance records, the surface defect incidence and degradation trends of various equipment types are extracted, such as the distribution of surface corrosion on transformers, crack history on circuit breaker casings, and loosening and corrosion records of tower fasteners. A neural network model comprehensively analyzes defect distribution characteristics, equipment operating time, and historical anomaly information to assign importance scores to inspection targets in the three-dimensional model and label their risks within the spatial model. This scoring mechanism breaks away from the traditional approach of optimizing inspection routes solely based on geographic paths. Instead, it dynamically plans inspection tasks based on surface defect risk priorities, better aligning resource allocation with the distribution of potential defects. During the inspection route planning phase, a reinforcement learning model's state space is constructed based on the three-dimensional model and importance scores. State variables include the drone's current location, power level, equipment importance score, and estimated flight cost. The action space is defined as a set of actions that select headings and flight distances based on different target equipment. The reward function comprehensively considers inspection time, number of defects detected, missed detection rate, and surface image quality metrics to optimize the strategy. The inspection strategy prioritizes high-importance equipment within a reasonable power range, such as aging transformers with severe surface corrosion, to avoid missed inspections of high-risk areas caused by the rigidity of conventional routes. During the inspection process, each drone performs autonomous flight according to the inspection planning strategy while simultaneously collecting visible light image streams of the target equipment in real time. A deep learning defect detection model is used to detect surface abnormalities such as cracks, rust, damage, and foreign matter attachment within the inspection image stream. During the inspection process, when the model identifies surface cracks or corroded areas, the system feeds the results back to the inspection task state space in real time and dynamically updates subsequent inspection actions based on the reinforcement learning model. For example, for transmission tower foot structures with identified surface abnormalities, the flight strategy is adjusted to include close-range hovering photography, capturing higher-resolution images to confirm the nature and extent of the defect, effectively reducing missed detection rates and improving inspection accuracy.This embodiment targets the needs of surface defect detection in power equipment. Through precise spatial modeling, importance scoring based on surface risks, reinforcement learning to optimize inspection paths, and dynamic adjustment of inspection linkage, it achieves intelligent, fine-grained, and highly responsive drone inspection tasks. This is significantly different from the existing inspection model that relies on fixed routes and has no real-time response, and effectively improves inspection efficiency, defect detection rate, and power equipment maintenance level.

[0064] Furthermore, the S1 includes the following steps:

[0065] Use several drones equipped with lidar and cameras to autonomously scan the power grid area, obtaining laser point cloud data and two-dimensional image sequences of the power grid area;

[0066] Performing filtering and registration processing based on the laser point cloud data to obtain a preliminary three-dimensional point cloud model of the power grid area;

[0067] The two-dimensional image sequence is subjected to multi-view image fusion processing to obtain a texture feature image of the power grid area, and the texture feature image is mapped to the preliminary three-dimensional point cloud model to obtain a three-dimensional model of the power grid area.

[0068] In some embodiments, the area to be inspected is first digitally reconstructed with high precision. Specifically, several multi-rotor drones equipped with laser radar (LiDAR) and high-resolution visible light cameras are deployed to conduct multiple autonomous flight scans of the target area. Flight missions are executed according to pre-defined geographic partitions and route planning, ensuring reasonable overlap between observation areas between different drones to improve modeling density and image fusion quality. During flight, each drone continuously acquires laser point cloud data and multi-angle two-dimensional image sequences within the target area. The point cloud data is used to describe the three-dimensional spatial structure of the surface of objects in the target space, while the image sequences are used for subsequent texture reconstruction. The acquired raw point cloud data contains a certain degree of noise and redundant point distribution due to factors such as laser reflectivity, flight attitude disturbances, and environmental occlusion. Therefore, the raw point cloud data is filtered, including outlier removal, ground separation, boundary thinning, and normal vector reconstruction. Subsequently, the point clouds from different drones are finely aligned in coordinate system using a multi-view point cloud registration algorithm (e.g., based on ICP or NDT), ultimately generating a preliminary three-dimensional point cloud model covering the entire power grid area structure. This model clearly depicts the spatial layout of equipment such as transformers, switchgear, disconnectors, bus corridors, and transmission towers. Furthermore, after performing image enhancement, dedistortion, and feature point extraction on the 2D image sequences captured by the drone, multi-view image fusion is achieved using methods such as structured beam adjustment and multi-view stereo (MVS) reconstruction to reconstruct a texture feature image with high spatial resolution. This image captures surface details such as rust on the transformer body, damaged insulators, and corroded connections in localized areas. After obtaining the texture feature image, a texture mapping algorithm (such as a triangulated mesh or voxel projection method) is used to fit the texture image to the surface of the preliminary 3D point cloud model with high fidelity. This ensures that every spatial equipment structural surface is characterized by visual features, ultimately generating a complete 3D textured power grid model. This model not only accurately restores the geometric structure but also preserves the surface texture details of the power equipment, providing precise spatial data for subsequent inspection object importance scoring, path planning, and surface defect identification. This modeling method is different from the existing technology that relies on a single drone and two-dimensional perception strategy based on planar images. It realizes high-precision three-dimensional spatial perception based on multi-source data fusion, and has significant advantages in accuracy improvement and information integrity.

[0069] Furthermore, the step S2 includes the following steps:

[0070] Based on the historical defect data of power equipment in the power grid area, a clustering algorithm is used to perform spatial hotspot analysis to obtain the initial hotspot data of the high-incidence areas of conventional defects in the power grid area;

[0071] Extract the locations of unrepaired defects and uninspected locations based on the last inspection result data to obtain the remaining defect data of the last inspection cycle;

[0072] spatially fusing the initial hotspot data with the defect legacy data to obtain a hotspot fusion data set;

[0073] Inputting the hotspot fusion data set into a convolutional neural network for training to obtain a scoring model for marking the importance score of the inspection object;

[0074] Based on the scoring model, a prediction scoring process is performed on the inspection objects in the three-dimensional model of the power grid area to obtain a three-dimensional power grid model marked with the importance scores of the inspection objects.

[0075] In some embodiments, a defect event feature vector, representing both spatial location and event frequency, is first constructed based on information such as the location, type, category, and time of occurrence of power equipment defects recorded in a historical defect database. This feature vector is then subjected to unsupervised spatial clustering using a density clustering algorithm. Specifically, a DBSCAN (Density-Based Spatial Clustering of Applications with Noise) method is used to identify high-density clusters of defect events in geographic space by setting a minimum sample number and spatial radius parameters. Sparse outliers are then removed, ultimately extracting a set of hotspots with high incidence of conventional defects within the power grid region, which are recorded as initial hotspot data. This initial hotspot data only reflects the statistical characteristics of fault distribution within the historical cycle and fails to capture potential risk factors remaining within the current cycle. To further enhance the model's ability to perceive the current task status, the locations of unrepaired defects and uninspected points recorded during the previous inspection are extracted. Unrepaired defects are identified by comparing historical fault labels with the current processing status using the device ID and geographic location. Uninspected points are extracted by comparing task scheduling logs with planned paths. The above two types of data are encoded into spatial point sets and assigned high-risk labels, and merged into the initial hotspot data to form a hotspot fusion data set containing structural historical clustering features and task residual factors. Subsequently, a data sample is constructed for training the scoring model. Taking each point in the fusion data set as the basic sample unit, its spatial coordinates, equipment type, historical defect intensity index (such as the average number of defects per year), the density of the number of adjacent defect points and the status label are extracted, and the corresponding supervision label is constructed to represent the inspection priority score of the point. After the above sample data is divided into a training set and a validation set, it is input into a convolutional neural network (CNN) model with a preset structure for training. The input layer of the CNN model accepts a two-dimensional or three-dimensional tensor structure inspection area coding map. The convolution layer uses multiple convolution kernels to extract spatial structural features and local risk distribution patterns. The pooling layer performs feature compression, and the fully connected layer generates the importance score of the corresponding inspection point at the output. The loss function uses the mean square error (MSE) or weighted regression error function for convergence optimization. After training, the scoring deviation and ranking consistency of the validation set results are verified to ensure that the model has good generalization ability and actual scoring reliability. Finally, the trained scoring model is deployed in the system's task scheduling module, where it performs scoring inference on each inspection object within the three-dimensional grid model. The system generates input feature vectors based on the inspection object's spatial location and equipment operating data. These vectors are fed into the CNN model, which then outputs corresponding importance scores and automatically annotates them to the corresponding object nodes within the three-dimensional grid model. The scoring results, expressed as heat values ​​or risk weights, serve as a crucial reference for subsequent path optimization and scheduling strategies.Compared with traditional strategies based on geographic location or manual experience-based weighted scoring, this method integrates historical cluster distribution with current legacy data and introduces a deep learning model to model risk spatial characteristics, thereby achieving risk identification based on efficient allocation of inspection resources.

[0076] Furthermore, the convolutional neural network is constructed by the following steps:

[0077] Based on the hotspot fusion dataset, the location coordinates, historical defect occurrence counts, most recent inspection defect detection results, and equipment type information of each inspection object are extracted to construct a training dataset containing inspection object spatial features, historical defect features, and type features;

[0078] Based on the training data set, the training data is divided into a training set and a validation set according to the importance labels corresponding to the inspection objects;

[0079] Inputting the training set into a convolutional neural network, performing feature extraction through a convolutional layer, and obtaining a multi-scale spatial feature map;

[0080] Performing pooling dimensionality reduction processing on the multi-scale spatial feature map to obtain a compressed feature vector;

[0081] Input the compressed feature vector into the fully connected layer, perform regression training, and output the importance score of the corresponding inspection object;

[0082] The prediction accuracy of the trained convolutional neural network is verified using the validation set.

[0083] In some embodiments, during the data preprocessing stage, the basic attribute information of each inspection object is extracted one by one from the fused data set, including the position coordinates (X, Y, Z) in three-dimensional space, the number of defects of the equipment in the historical records, the output results of the defect detection model during the most recent inspection (including confidence, defect type), and the equipment type code (such as transformer, circuit breaker, disconnector, line support, etc.). The above data together constitute a multi-dimensional input feature vector of the inspection object. In order to improve the generalization ability and training stability of the model, the above feature vectors are standardized so that the features of each dimension are in a uniform numerical range. Then, based on the importance label corresponding to each inspection object in the fused data set (generated by expert experience or inspection history scoring system), the overall data is divided into a training set and a validation set, usually in a ratio of 8:2 or 7:3, to ensure that the training set fully covers the feature distribution of various samples. The label value is expressed in the form of a continuous value, which is suitable for regression task modeling in neural networks. In terms of model design, the constructed convolutional neural network consists of multiple layers of stacked convolutional modules. The input layer receives a feature map of inspection objects, which is mapped into a tensor structure after vector re-arrangement. The convolutional layer uses convolution kernels of different sizes to extract the interaction between local spatial adjacency and multi-dimensional attributes, forming a multi-scale spatial feature map. This layer captures the underlying high-order patterns of inspection objects in the spatial, temporal, and device attribute dimensions. After feature extraction, a pooling layer performs dimensionality reduction to compress redundant information and enhance the model's translation invariance. Maxpooling or average pooling is often used. The pooling output serves as a deep semantic feature and is further input into a fully connected layer for feature fusion and regression modeling, ultimately outputting an importance score for each inspection object. During training, the mean squared error (MSE) is used as the loss function to measure the deviation between the network's predicted scores and the true labels. The optimizer uses the Adam or RMSProp algorithm to improve gradient convergence efficiency. To prevent overfitting, the Dropout regularization strategy is introduced, and the early stopping mechanism is used to monitor the error changes of the validation set and dynamically terminate the training process. After the model training is completed, its prediction accuracy and ranking consistency are evaluated on the validation set. The verification indicators include mean absolute error, determination coefficient (R 2) and the Spearman rank correlation coefficient are used to measure the robustness and effectiveness of the model in scoring continuity and relative ranking. The final convolutional neural network scoring model can accept the input features of the equipment to be evaluated in any power grid area during the inference stage and output the corresponding importance scoring results. The scoring results will serve as an important reference factor for the input of subsequent path planning algorithms to achieve risk-based drone inspection priority sorting. Compared with traditional scoring methods that rely on static rule settings or manual experience grading, this solution has self-learning capabilities and can adapt to complex power equipment structures and defect evolution characteristics, significantly improving the intelligence level of inspection strategies and the dynamic response capabilities of scoring mechanisms.

[0084] Furthermore, the step S3 includes the following steps:

[0085] Based on the three-dimensional model of the power grid area, the spatial location and corresponding importance score of each inspection object are extracted. The inspection task state space is constructed by combining the initial position, power state, flight speed and defect confirmation status of each drone.

[0086] Based on the inspection task state space, an action space is defined as an action set for each UAV to select a target inspection object in the next step, and the action set includes action options of heading to the inspection object with different headings and different ranges;

[0087] The inspection task state space and the action space are input into the reinforcement learning model, and the next inspection target point of each drone is output through action decision-making to obtain a preliminary drone inspection path sequence;

[0088] Based on the preliminary UAV inspection path sequence, the inspection path strategy is optimized through a reward function. The reward value of the reward function is calculated based on the inspection time, the number of defect targets found, the missed detection rate, and the clarity of the defect image;

[0089] Based on the strategy optimization results, the reinforcement learning model parameters are updated and a real-time drone inspection route planning strategy is generated.

[0090] In some embodiments, a state space for the reinforcement learning task is first constructed. This state space is based on each drone's current flight state and the global task distribution. Specifically, it includes: each drone's spatial position, remaining battery level, flight speed, target reachability indicator at the current time step, as well as the three-dimensional coordinates, importance scores, and defect confirmation status (e.g., whether confirmed, remaining high-risk targets, etc.) of all inspection objects. These state parameters collectively constitute a time-series state vector, which serves as the input to the policy network. Regarding action space design, a set of optional actions is defined for each drone. The elements of the set are the inspection target points it may choose to visit next and the corresponding flight strategy combinations. The action set not only considers the target point location and whether the remaining battery level supports round trips, but also incorporates the current heading offset angle and flight cost to form a discrete action encoding structure. Typically, candidate actions include "travel to the point with the highest importance score," "travel to the nearest uninspected target," and "execute hover wait." To enhance model convergence efficiency, the size of the action space is limited during training using a top-N pruning strategy among candidate actions to reduce search complexity. During policy learning, the aforementioned state and action spaces are used as inputs, and policy optimization is performed using a policy gradient method based on a deep reinforcement learning framework. An actor-critic architecture is employed. The policy network (Actor) accepts a state vector and outputs an action probability distribution. Based on this distribution, it samples inspection actions for the current time step. The value function network (Critic) evaluates the value of state-action pairs, which guides policy gradient updates. The evaluation of inspection path sequences relies on a reward function, which calculates the corresponding cumulative reward after each path execution. The reward function designed in this scheme fully considers the practical requirements of multi-objective optimization for inspection tasks. It constructs a comprehensive reward using a weighted combination of the following sub-indicators: First, inspection task completion efficiency, which is defined as the number of targets completed per unit time and is defined as a negative function of inspection time; second, defect detection capability, with a positive reward calculated based on the number of defect targets discovered per unit time; third, task coverage completeness, measured as the complement of the missed detection rate (1-missed detection rate); and fourth, image quality evaluation, which combines image clarity scores (e.g., based on structural similarity (SSIM) or mean resolution) to score images captured near defect areas, encouraging high-quality image acquisition. The final reward function is expressed as a weighted linear combination of these sub-indicators and employs regularization to stabilize the training process. The reinforcement learning model employs an experience replay mechanism and a target network structure during training to enhance training sample diversity and convergence stability. After training, the optimal policy model is deployed to the task scheduling system. In real-time, it receives state input and outputs the next target action for each drone, dynamically generating a preliminary inspection path sequence.During operation, each time the current action is completed and defect detection results are obtained, the state space is updated in real time. Based on the new inspection image and target state feedback information, the policy network is called to make the next round of action decisions, forming a dynamic path optimization closed loop with time steps as the unit. Compared with traditional inspection systems based on fixed path tables, the reinforcement learning path planning method described in this embodiment has significant adaptability and task sensitivity. It can dynamically optimize resource allocation based on equipment risk level, task execution status, and defect feedback information, maximizing inspection efficiency and defect detection quality, and has strong algorithmic innovation and application effectiveness.

[0091] Furthermore, the reward function is as follows:

[0092]

[0093] Among them, R is the reward value; T is the cumulative time it takes for the drone to complete the inspection task; T ref is the preset reference time; D is the number of defect targets successfully found during the inspection process; L is the missed detection rate of the inspection task; Q is the clarity score of the defect image collected close to the defect target; Q max is the maximum reference value of the clarity score; α1, α2 and α4 are the reward coefficients; α3 is the penalty coefficient.

[0094] It's important to note that the reward function first measures the timeliness of task execution based on the ratio between the actual time a drone takes to complete an inspection task and a system-set reference time. When the actual time is lower than the reference value, a positive reward is awarded; otherwise, a negative reward is incurred, guiding the policy network to prioritize inspection sequences with shorter paths and faster execution. Secondly, a metric representing the number of defect targets discovered is incorporated into the reward function to positively incentivize the model to focus inspections on areas with high defect risk or high-scoring equipment, thereby improving the overall defect detection rate. Furthermore, to account for the potential for missed inspections due to insufficient path coverage, the reward function further incorporates a complement to the missed detection rate, reflecting the completeness of inspection coverage in the form of a task completion rate. This metric guides the model to maximize global coverage while balancing time efficiency with risk prioritization, preventing it from focusing excessively on high-scoring areas and neglecting other targets. Furthermore, to ensure the readability and defect identification value of inspection images, the reward function also includes an image clarity score, which reflects the quality of images captured by the drone when approaching the equipment surface. Image quality metrics can be based on the output of an image clarity assessment algorithm, such as the Structural Similarity Index (SSIM) or gradient strength, to reflect the clarity of image detail edges and the ability to identify defect features. This incentive mechanism encourages the strategy to adopt appropriate flight angles and altitudes to acquire high-quality images when approaching high-risk targets. Overall, the reward function is a weighted combination of multiple sub-indicators, with coefficients adjusted based on mission objectives to achieve optimal policy balance under multi-objective tasks. Through the gradual feedback and optimization process of this reward function, the reinforcement learning model adaptively learns and generates inspection routing strategies with high execution efficiency, high defect coverage, and guaranteed image acquisition quality. Compared to traditional strategies that focus solely on optimizing the shortest path or the number of defect points, this reward mechanism achieves more comprehensive judgment capabilities and practical engineering applicability for intelligent scheduling.

[0095] Furthermore, the S4 includes the following steps:

[0096] S41, based on the inspection route planning strategy generated in real time, extracting the inspection action sequence of each drone, and generating an inspection task instruction set including the inspection target location, action instructions and execution order;

[0097] S42. Based on the inspection task instruction set, each UAV is dispatched and controlled to make the UAV fly autonomously according to the action sequence and synchronously collect flight status data and inspection image stream data at each time step;

[0098] S43, inputting the inspection image stream data into a defect detection model to perform real-time defect detection, obtaining a defect detection result of the current time step, and updating the inspection task state space based on the defect detection result;

[0099] S44. Based on the updated inspection task state space, re-input the reinforcement learning model, generate an inspection route planning strategy, and update the inspection task instruction set; if the inspection task instruction set is not empty, execute S42; if the inspection task instruction set is empty, execute S45;

[0100] S45. According to the inspection defect detection results and the inspection target location information, the defect location is marked in the three-dimensional model of the power grid area, and the defect image, equipment number and detection time are associated and stored to generate an inspection data record.

[0101] In some embodiments, the inspection action sequence required for each drone in the current mission round is first extracted from the path planning strategy output by the reinforcement learning model. This action sequence includes several action instructions with spatial target identifiers. Each instruction clearly corresponds to the spatial position code of the target device, the heading angle required for drone flight, the preset flight speed, the hovering time during shooting, and the sequential logic of task execution, thereby forming a structured inspection task instruction set to ensure the timing and trajectory controllability of task execution. The system then calls the flight control interface to dispatch and control each drone according to the inspection task instruction set. When executing each action instruction, the drone must complete the approach flight and attitude adjustment to the designated target device with the support of precise positioning and trajectory planning, and collect multi-view images within the preset hovering time. To ensure data closed-loop capability, the drone synchronously collects its own flight status data (such as GPS coordinates, power, speed, attitude angle, etc.) and inspection image stream data in the forward and downward directions at each time step. The image data is transmitted back to the edge processing unit or cloud inference server in real time in the form of video frames or continuous static images. The collected image stream data is immediately fed into the deployed defect detection model in the edge computing unit for identification and processing. Based on a deep convolutional neural network architecture, this model extracts features and classifies surface defects (such as rust, cracks, and insulation damage) on power equipment. It outputs defect detection results for each target device at the current time step, including defect type, location bounding box, and confidence score. These detection results not only serve as the basis for fault determination but also provide feedback for dynamically updating the inspection task state space. Specifically, when the system identifies a device with clear signs of a defect that has not yet been confirmed and addressed, the device node status is updated from "pending inspection" to "marked" and assigned a high inspection priority label. If the current target is confirmed to be free of anomalies, the corresponding node status is changed to "completed" to reduce subsequent rescheduling. The updated task state space is re-input into the reinforcement learning path planning module, triggering a new round of policy generation. The reinforcement learning model dynamically adjusts the path planning logic based on the current flight status, the distribution of uncompleted task points, and the image recognition results, and generates a new set of inspection task instructions, enabling adaptive task scheduling throughout the entire process. The system determines whether there are still pending inspection tasks based on the instruction set. If the instruction set is not empty, it loops back to the dispatch control module to continue the inspection. If the instruction set is empty, it assumes that all tasks in the current round have been completed, and the system enters the post-processing phase. After the task is completed, the system uses the final defect detection results and equipment spatial information to perform layer-level annotation of the target locations of defects in the three-dimensional model of the power grid area, generating a defect distribution map with spatial positioning information. Each defect information is simultaneously bound to its corresponding image data, equipment number, detection time, detection model output confidence level, and other parameters, and uniformly recorded in the inspection data recording system.The above records can be used for subsequent fault repair scheduling, equipment operation and maintenance analysis, and risk prediction model training.

[0102] Furthermore, the defect detection model is deployed on a cloud server that is communicatively connected to the drone.

[0103] It should be noted that to balance model accuracy and real-time performance requirements for surface defect identification on power equipment, the defect detection model is deployed on a cloud server with high-performance computing capabilities. Data transmission and inference results are exchanged via a low-latency communication link with the drone system. This model utilizes a deep convolutional neural network architecture with multi-scale texture perception and location boundary recognition capabilities. It is specifically designed to identify common surface defects such as cracks, rust, damage, insulator contamination, and loose conductors on power grid equipment. During inspections, the drone uses an edge module to pre-process the image stream (e.g., resizing, frame extraction, and noise filtering). The processed image information is then uploaded in real time to a cloud-based inspection platform in the form of compressed data packets via wireless communication channels such as 5G or LTE. Upon receiving the image data, the cloud server immediately invokes the deployed defect detection model for online inference, outputting structured detection results including defect type, location coordinates, confidence score, and image mask of the suspected target area. This detection process is performed in the cloud using batch inference or parallel GPU acceleration, ensuring low latency and high throughput for large-scale image stream processing.

[0104] Furthermore, the defect detection model is constructed by the following steps:

[0105] Based on the historical image data of power equipment inspections in the power grid, original image samples are extracted and annotated to obtain training and verification datasets containing defect target categories and location box annotations;

[0106] Based on the training data set, performing size normalization, color normalization and data enhancement processing on the image samples to obtain preprocessed training image data;

[0107] Build the Transformer network model and initialize the network parameters;

[0108] Inputting the preprocessed training image data into the Transformer network model, and optimizing the model parameters by jointly training with weighted cross entropy loss and bounding box regression loss;

[0109] The detection accuracy and positioning accuracy of the Transformer network model are evaluated based on the validation dataset, and the model hyperparameters are adjusted according to the evaluation results to complete the defect detection model construction.

[0110] Specifically, based on a database of historical images from power equipment inspections, raw image samples covering various equipment surface conditions were screened, including typical defect examples such as transformer casing cracks, circuit breaker contact surface corrosion, loose transmission line clamps, insulator contamination spots, and lightning arrester damage. Fine-grained manual annotation of each extracted raw image was performed using a semi-automatic annotation platform. The annotations included the defect target's category label and corresponding location information, and the defect area was marked with a bounding box. Finally, a training dataset and a validation dataset were constructed, containing image data, category labels, and location box annotations. During the data preprocessing phase, image samples in the training dataset were normalized to improve model training stability and generalization. During size normalization, all images were scaled and padded to a fixed input resolution (e.g., 512×512 or 640×640 pixels) to accommodate the fixed input size requirement of the Transformer model's self-attention mechanism. Color normalization unified the mean and standard deviation of each channel to eliminate data drift caused by varying lighting conditions. To further enhance the model's robustness to defect signatures at varying scales, angles, and lighting conditions, various data augmentation strategies, including random horizontal flipping, random affine transformation, color jittering, and random cropping, are implemented to expand the training sample space. The model's main architecture utilizes an object detection model based on the Transformer encoder-decoder architecture. Network parameters are first initialized, including the weight matrices of the encoder's self-attention layer and feedforward neural network module, and the decoder's query vector and output head weights. Appropriate parameter normalization is then performed to enhance initial training stability. Preprocessed training images are fed into the Transformer network. Initial feature maps are extracted from the images through a shallow convolutional network. These images are then fed into the encoder for multi-layer, multi-head self-attention feature modeling, capturing long-range dependencies and local texture features. The decoder predicts defect instances in the image based on a set of learnable query vectors and outputs a class probability distribution and corresponding bounding box coordinate parameters for each object. During training, the optimization objective is to jointly minimize the classification and localization losses. The classification loss uses a weighted cross-entropy function to address the uneven distribution of defect classes, assigning higher weights to rare defect classes to improve detection accuracy for small sample sizes. The localization loss uses a bounding box regression loss, such as the GIoU or DIoU loss, to optimize the positional match between the predicted and ground-truth boxes. The overall loss function is a weighted sum of the classification and regression losses, and the model parameters are iteratively optimized using a backpropagation algorithm. After training, the model performance is evaluated using an independent validation dataset. Evaluation metrics include defect detection accuracy (Precision), recall (Recall), mean average precision (mAP), and localization error (mean Intersection over Union).Based on the performance feedback from the validation set, the model's hyperparameters were fine-tuned, including adjusting the learning rate, optimizing the regularization coefficient, and selecting the number of Transformer encoder layers, to further improve the final model's detection stability and positioning accuracy in actual inspection environments. This complete process ultimately led to the construction of a Transformer deep learning model suitable for surface defect detection in power equipment. This model possesses the ability to adaptively recognize a variety of complex surface defect features, significantly outperforming traditional convolutional networks in long-range feature modeling and multi-target detection accuracy.

[0111] The present invention also includes a power equipment defect detection system based on deep learning, comprising:

[0112] The power grid model module is used to build a three-dimensional model of the power grid area;

[0113] The inspection marking module is used to mark the importance scores of inspection objects on the three-dimensional model of the power grid area through a neural network model based on the historical defect data of power equipment in the power grid area;

[0114] An intelligent planning module, which uses a reinforcement learning model to build a UAV inspection route planning strategy based on the importance scores of the inspection objects and the three-dimensional model of the power grid area;

[0115] The defect detection module is used to schedule drone inspections according to the drone inspection route planning strategy and simultaneously obtain inspection image streams for power equipment defect detection.

[0116] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A method for detecting defects in power equipment based on deep learning, characterized in that: include: S1. Construct a three-dimensional model of the power grid area; S2. Based on the historical defect data of power equipment in the power grid area, a neural network model is used to mark the importance scores of inspection objects in the three-dimensional model of the power grid area; S3. Based on the three-dimensional model of the power grid area, a UAV inspection route planning strategy is constructed through a reinforcement learning model; S4. Schedule the drone for inspection according to the drone inspection route planning strategy, and simultaneously obtain the inspection image stream to detect defects in power equipment.

2. The method for detecting defects in power equipment based on deep learning according to claim 1, characterized in that: Said S1 comprises the following steps: Use several drones equipped with lidar and cameras to autonomously scan the power grid area, obtaining laser point cloud data and two-dimensional image sequences of the power grid area; Performing filtering and registration processing based on the laser point cloud data to obtain a preliminary three-dimensional point cloud model of the power grid area; The two-dimensional image sequence is subjected to multi-view image fusion processing to obtain a texture feature image of the power grid area, and the texture feature image is mapped to the preliminary three-dimensional point cloud model to obtain a three-dimensional model of the power grid area.

3. The method for detecting defects in power equipment based on deep learning according to claim 1, characterized in that: The S2 comprises the following steps: Based on the historical defect data of power equipment in the power grid area, a clustering algorithm is used to perform spatial hotspot analysis to obtain the initial hotspot data of the high-incidence areas of conventional defects in the power grid area; Extract the locations of unrepaired defects and uninspected locations based on the last inspection result data to obtain the remaining defect data of the last inspection cycle; spatially fusing the initial hotspot data with the defect legacy data to obtain a hotspot fusion data set; Inputting the hotspot fusion data set into a convolutional neural network for training to obtain a scoring model for marking the importance score of the inspection object; Based on the scoring model, a prediction scoring process is performed on the inspection objects in the three-dimensional model of the power grid area to obtain a three-dimensional power grid model marked with the importance scores of the inspection objects.

4. The method for detecting defects in power equipment based on deep learning according to claim 3, characterized in that: The convolutional neural network is constructed by the following steps: Based on the hotspot fusion dataset, the location coordinates, historical defect occurrence counts, most recent inspection defect detection results, and equipment type information of each inspection object are extracted to construct a training dataset containing inspection object spatial features, historical defect features, and type features; Based on the training data set, the training data is divided into a training set and a validation set according to the importance labels corresponding to the inspection objects; Inputting the training set into a convolutional neural network, performing feature extraction through a convolutional layer, and obtaining a multi-scale spatial feature map; Performing pooling dimensionality reduction processing on the multi-scale spatial feature map to obtain a compressed feature vector; Input the compressed feature vector into the fully connected layer, perform regression training, and output the importance score of the corresponding inspection object; The prediction accuracy of the trained convolutional neural network is verified using the validation set.

5. The method for detecting defects in power equipment based on deep learning according to claim 1, characterized in that: The S3 includes the following steps: Based on the three-dimensional model of the power grid area, the spatial location and corresponding importance score of each inspection object are extracted. The inspection task state space is constructed by combining the initial position, power state, flight speed and defect confirmation status of each drone. Based on the inspection task state space, an action space is defined as an action set for each UAV to select a target inspection object in the next step, and the action set includes action options of heading to the inspection object with different headings and different ranges; The inspection task state space and the action space are input into the reinforcement learning model, and the next inspection target point of each drone is output through action decision-making to obtain a preliminary drone inspection path sequence; Based on the preliminary UAV inspection path sequence, the inspection path strategy is optimized through a reward function. The reward value of the reward function is calculated based on the inspection time, the number of defect targets found, the missed detection rate, and the clarity of the defect image; Based on the strategy optimization results, the reinforcement learning model parameters are updated and a real-time drone inspection route planning strategy is generated.

6. The method for detecting defects in power equipment based on deep learning according to claim 5, characterized in that: The reward function is as follows: Among them, R is the reward value; T is the cumulative time it takes for the drone to complete the inspection task; T ref is the preset reference time; D is the number of defect targets successfully found during the inspection process; L is the missed detection rate of the inspection task; Q is the clarity score of the defect image collected close to the defect target; Q max is the maximum reference value of the clarity score; α1, α2 and α4 are the reward coefficients; α3 is the penalty coefficient.

7. The method for detecting defects in power equipment based on deep learning according to claim 5, characterized in that: The S4 comprises the following steps: S41, based on the inspection route planning strategy generated in real time, extracting the inspection action sequence of each drone, and generating an inspection task instruction set including the inspection target location, action instructions and execution order; S42. Based on the inspection task instruction set, each UAV is dispatched and controlled to make the UAV fly autonomously according to the action sequence and synchronously collect flight status data and inspection image stream data at each time step; S43, inputting the inspection image stream data into a defect detection model to perform real-time defect detection, obtaining a defect detection result of the current time step, and updating the inspection task state space based on the defect detection result; S44. Based on the updated inspection task state space, re-input the reinforcement learning model, generate an inspection route planning strategy, and update the inspection task instruction set; if the inspection task instruction set is not empty, execute S42; if the inspection task instruction set is empty, execute S45; S45. According to the inspection defect detection results and the inspection target location information, the defect location is marked in the three-dimensional model of the power grid area, and the defect image, equipment number and detection time are associated and stored to generate an inspection data record.

8. The method for detecting defects in power equipment based on deep learning according to claim 7, characterized in that: The defect detection model is deployed on a cloud server that is connected to the drone.

9. The method for detecting defects in power equipment based on deep learning according to claim 8, characterized in that: The defect detection model is constructed by the following steps: Based on the historical image data of power equipment inspections in the power grid, original image samples are extracted and annotated to obtain training and verification datasets containing defect target categories and location box annotations; Based on the training data set, performing size normalization, color normalization and data enhancement processing on the image samples to obtain preprocessed training image data; Build the Transformer network model and initialize the network parameters; Inputting the preprocessed training image data into the Transformer network model, and optimizing the model parameters by jointly training with weighted cross entropy loss and bounding box regression loss; The detection accuracy and positioning accuracy of the Transformer network model are evaluated based on the validation dataset, and the model hyperparameters are adjusted according to the evaluation results to complete the defect detection model construction.

10. A power equipment defect detection system based on deep learning, applied to a power equipment defect detection method based on deep learning according to any one of claims 1 to 9, characterized in that: include: The power grid model module is used to build a three-dimensional model of the power grid area; The inspection marking module is used to mark the importance scores of inspection objects on the three-dimensional model of the power grid area through a neural network model based on the historical defect data of power equipment in the power grid area; An intelligent planning module, which uses a reinforcement learning model to build a UAV inspection route planning strategy based on the importance scores of the inspection objects and the three-dimensional model of the power grid area; The defect detection module is used to schedule drone inspections according to the drone inspection route planning strategy and simultaneously obtain inspection image streams for power equipment defect detection.

Images