A deep learning-based power equipment defect detection system and method

Abstract

The present application 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, which constructs a three-dimensional model of a power grid area, and based on historical defect data, uses a neural network model to mark the importance score of the inspection object in the three-dimensional model, completes the quantitative evaluation of the equipment fault risk level, and improves the matching degree of the inspection resource and the defect risk distribution. Combined with the three-dimensional model of the power grid and the importance score, an unmanned aerial vehicle inspection route planning strategy is constructed using a reinforcement learning model, so that the inspection path comprehensively considers the power equipment defect risk and the spatial factors in the planning stage, optimizes the task allocation, and overcomes the problem that the static path cannot adapt to the changes of the equipment. In the inspection execution process, the unmanned aerial vehicle is dispatched according to the planning strategy, and the image stream is synchronously acquired for defect detection, realizing the linkage of the inspection action and the detection process, and improving the response efficiency and detection quality for potential defects.

Description

TECHNICAL FIELD

[0001] The present application 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

[0002] The existing unmanned aerial vehicle (UAV) inspection mostly adopts a long-distance patrol mode, that is, continuous cruising along a preset power line, substation boundary or fixed flight route, and lacks targeted risk assessment and priority arrangement for the inspection object. This long-distance patrol mode usually plans the flight route based on the principle of geographical continuity or shortest path, ignoring the significant differences in the running state, historical defect occurrence probability and safety importance of different power equipment. Since the importance of the inspection object is not fully considered, the investment of inspection resources does not match the fault discovery benefit, resulting in waste of flight resources and reduced detection effect.

[0003] Secondly, the existing UAV inspection path is generally based on static flight route planning, which usually determines the flight route and task points once before the start of the inspection task, and does not have the ability to dynamically optimize the inspection path according to the real-time detection state of the power equipment during the flight. The current inspection system only supports image acquisition or observation tasks at preset waypoints, and lacks linkage with real-time defect detection systems. When the UAV discovers new potential defect targets during the inspection process, it cannot adjust the remaining inspection plan according to real-time feedback, such as increasing close-range flight or hovering observation of suspected defect areas, which may lead to missed detection or detection failure. In addition, since the inspection path cannot be flexibly adjusted according to the change of equipment state, the UAV may continue to fly along the original route even if new abnormal signals are encountered, which affects the quality of power equipment defect detection in the power grid. SUMMARY

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

[0005] To achieve the above purpose, the technical solution adopted by the present application is:

[0006] A power equipment defect detection method based on deep learning, comprising:

[0007] S1, constructing a three-dimensional model of a power grid area;

[0008] S2, based on historical defect data of power equipment in the power grid area, marking the importance score of the inspection object in the three-dimensional model of the power grid area through a neural network model;

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

[0010] S4, scheduling the unmanned aerial vehicle to perform inspection according to the unmanned aerial vehicle inspection route planning strategy, and synchronously acquiring an inspection image stream to perform power equipment defect detection.

[0011] Further, the S1 comprises the following steps:

[0012] A plurality of unmanned aerial vehicles carrying laser radars and cameras are used to autonomously scan the power grid area to obtain laser point cloud data and a two-dimensional image sequence of the power grid area.

[0013] The laser point cloud data is filtered and registered 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 into the preliminary three-dimensional point cloud model to obtain a three-dimensional model of the power grid area.

[0015] Further, the S2 comprises the following steps:

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

[0017] According to the last inspection result data, un-repaired defect positions and un-inspected positions are extracted to obtain defect remaining data of the last inspection cycle.

[0018] The initial hotspot data and the defect remaining data are spatially fused to obtain a hotspot fusion data set.

[0019] The hotspot fusion data set is input 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, the inspection object in the three-dimensional model of the power grid area is subjected to predictive scoring processing to obtain a three-dimensional power grid model marked with the importance score of the inspection object.

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

[0022] Based on the hotspot fusion data set, the position coordinates, the number of historical defect occurrences, the last inspection defect detection result, and the equipment type information of each inspection object are extracted to construct a training data set containing the spatial features, historical defect features, and type features of the inspection object.

[0023] Based on the training data set, the training data is divided into a training set and a validation set according to the importance label corresponding to the inspection object.

[0024] Input the training set into a convolutional neural network, perform feature extraction through a convolutional layer, and obtain a multi-scale spatial feature map;

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

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

[0027] Verify the prediction accuracy of the trained convolutional neural network through the validation set.

[0028] Further, the S3 comprises the following steps:

[0029] Based on the three-dimensional model of the power grid region, the spatial position of each inspection object and the corresponding importance score are extracted, and the initial position, power state, flight speed of each unmanned aerial vehicle and the defect confirmation state of the inspection object are combined to construct an inspection task state space;

[0030] Based on the inspection task state space, define the action space as the action set of each unmanned aerial vehicle selecting a target inspection object in the next step, and the action set includes action options for approaching the inspection object at different headings and different distances;

[0031] Input the inspection task state space and the action space into a reinforcement learning model, output the next inspection target point of each unmanned aerial vehicle through action decision, and obtain a preliminary unmanned aerial vehicle inspection path sequence;

[0032] Based on the preliminary unmanned aerial vehicle inspection path sequence, perform policy optimization on the inspection path through a reward function, and the reward value of the reward function is calculated according to the inspection time, the number of discovered defect targets, the missed detection rate and the defect image clarity;

[0033] According to the policy optimization result, update the reinforcement learning model parameters and generate a real-time unmanned aerial vehicle inspection route planning strategy.

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

[0035]

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

[0037] Further, the S4 comprises the following steps:

[0038] S41, based on the real-time generated inspection route planning strategy, extracting the inspection action sequence of each unmanned aerial vehicle, and generating an inspection task instruction set containing an inspection target position, an action instruction, and an execution sequence;

[0039] S42, based on the inspection task instruction set, scheduling and controlling each unmanned aerial vehicle, so that the unmanned aerial vehicle autonomously flies according to the action sequence, and synchronously collects flight state data and inspection image stream data at each time step;

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

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

[0042] S45, according to the inspection defect detection result and the inspection target position information, performing defect position labeling in a three-dimensional model of the power grid area, and associating and storing the defect image, the equipment number, and the detection time to generate an inspection data record.

[0043] Further, the defect detection model is deployed on a cloud server in communication connection with the unmanned aerial vehicle.

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

[0045] Based on the inspection historical image data of the power equipment in the power grid, extracting original image samples, and labeling the image samples to obtain a training data set and a verification data set containing defect target categories and position box labels;

[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] Constructing a Transformer network model and initializing network parameters;

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

[0049] Evaluate the detection accuracy and positioning accuracy of the Transformer network model based on the verification data set, and adjust the model hyperparameters according to the evaluation results to complete the construction of the defect detection model.

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

[0051] A power grid model module is used to construct a three-dimensional model of the power grid area.

[0052] An inspection marking module is used to mark the importance score of the inspection object in the three-dimensional model of the power grid area based on the historical defect data of the power equipment in the power grid area through a neural network model.

[0053] An intelligent planning module is used to construct a UAV inspection route planning strategy based on the importance score of the inspection object and the three-dimensional model of the power grid area through a reinforcement learning model.

[0054] A defect detection module is used to schedule the UAV inspection according to the UAV inspection route planning strategy, and simultaneously acquire the inspection image stream for power equipment defect detection.

[0055] The application has the beneficial effects that: the application realizes accurate modeling of the spatial distribution of power equipment by constructing a three-dimensional model of the power grid area, avoiding the resource waste and blind area problems caused by the existing route planning mode mainly based on line continuity. Based on the historical defect data of the power equipment in the power grid area, the importance score of the inspection object is marked in the three-dimensional model of the power grid area by using a neural network model, the quantitative evaluation and classification of the equipment fault risk level are completed, and the problem of lack of task priority sorting based on equipment operation risk in the existing inspection process is solved, and the matching degree of inspection resources and defect risk distribution is improved. Based on the three-dimensional model of the power grid area and the importance score information of the inspection object, an unmanned aerial vehicle inspection route planning strategy is constructed by using a reinforcement learning model, so that the inspection path can comprehensively consider factors such as equipment importance, spatial position and flight cost in the planning stage, and the inspection task allocation and path selection are optimized, and the limitations of the existing static fixed inspection path and the inability to adapt to changes in power equipment are overcome. In the inspection execution stage, the unmanned aerial vehicle is dispatched according to the unmanned aerial vehicle inspection route planning strategy, and the inspection image stream is synchronously acquired to execute power equipment defect detection, realize the linkage of the inspection action and the defect detection process, improve the real-time response capability to potential defects, effectively reduce the omission rate, improve the attention of the unmanned aerial vehicle to the defect problem of the power equipment, and realize higher image quality for the defect problem to improve the detection accuracy. Through the above technical means, the dynamic identification of the importance of the inspection object, the intelligent optimization of the inspection path and the linkage control of the inspection action in the power equipment inspection process are effectively realized, the inspection efficiency, the defect discovery rate and the detection quality are significantly improved, and the deficiencies of the existing unmanned aerial vehicle inspection system in task allocation, path planning and defect detection response are overcome. BRIEF DESCRIPTION OF DRAWINGS

[0056] Fig. 1 is a step flow chart of a power equipment defect detection method based on deep learning in the application.

[0057] Fig. 2 is a step flow chart of step S4 in the application. DETAILED DESCRIPTION

[0058] Please refer to Figs. 1-2 The application relates to a power equipment defect detection method based on deep learning, which comprises the following steps:

[0059] S1, constructing a three-dimensional model of a power grid area;

[0060] S2, based on historical defect data of power equipment in the power grid area, marking the importance score of the inspection object in the three-dimensional model of the power grid area by using a neural network model;

[0061] S3, based on the three-dimensional model of the power grid area, constructing an unmanned aerial vehicle inspection route planning strategy by using a reinforcement learning model;

[0062] S4, scheduling the unmanned aerial vehicle according to the unmanned aerial vehicle inspection route planning strategy, and synchronously acquiring an inspection image stream to detect defects of the power equipment.

[0063] It should be noted that for the surface defects that may exist on the surface of power equipment such as transformers, circuit breakers, disconnectors, transmission line towers and arresters in the power grid area, including cracks, corrosion, rust, damage, insulation contamination and foreign matter attachment, etc. Surface abnormalities, the present embodiment first scans the power grid area autonomously by multiple unmanned aerial vehicles equipped with laser radars and high-resolution visible light cameras. For key detection parts such as transformer tank, circuit breaker contact, disconnector connecting rod, transmission line tower structure and arrester shell, fine point cloud data and surface texture images are collected. Through point cloud filtering, registration and texture fusion, a high-precision power grid three-dimensional model is generated, which completely restores the surface characteristics of various power equipment, providing a spatial basis for subsequent defect detection and inspection path planning. Compared with the traditional route patrol method, this scheme can accurately cover the parts with complex surface structure of the equipment, which are prone to crack corrosion and other defects, significantly improving the inspection coverage rate and modeling accuracy. Based on the completed power grid three-dimensional model, combined with historical maintenance records, the surface defect occurrence rate and deterioration trend of various equipment are extracted, such as transformer surface rust distribution, circuit breaker shell crack history, tower fastener loosening and corrosion record, etc. Through neural network model comprehensive analysis of defect distribution characteristics, equipment running time and historical abnormal information, the importance score of the inspection object in the three-dimensional model is calculated, and the risk is labeled in the spatial model. This scoring mechanism breaks through the traditional method of relying only on geographical path optimization of inspection route, dynamically plans the inspection task based on the risk priority of surface defects, making the resource allocation and potential defect distribution more matched. In the inspection path planning stage, based on the three-dimensional model and importance score, the state space of the reinforcement learning model is constructed, and the state variables include the current position of the unmanned aerial vehicle, the power state, the importance score of the equipment and the estimated flight cost. The action space is defined as the action set based on the selection of heading and flight distance for different target equipment, and the reward function considers the inspection time, the number of detected defects, the miss detection rate and the surface image quality index for strategy optimization. The inspection strategy preferentially guides the unmanned aerial vehicle to inspect the equipment with high importance score within a reasonable power range, such as the aged transformer with severe surface corrosion, to avoid the situation of missing detection of high-risk parts caused by fixed conventional path. During the inspection execution process, each unmanned aerial vehicle executes autonomous flight according to the inspection planning strategy, while real-time visible light image stream of the target equipment is collected. For the inspection image stream, a deep learning defect detection model is used to detect surface crack, rust, damage, foreign matter attachment and other appearance abnormal features. During the detection process, when the model identifies a surface crack or corrosion area, the system feeds back the detection result to the inspection task state space in real time, and dynamically updates the subsequent inspection action based on the reinforcement learning model, such as adjusting the flight strategy to increase the close-range hovering for the transmission tower foot structure with identified surface abnormalities, to capture higher resolution images to confirm the defect nature and range, thereby effectively reducing the miss detection rate and improving the detection accuracy.The embodiment aims at the surface defect detection requirement of power equipment, realizes the intelligentization, fine granularity and high responsiveness of the unmanned aerial vehicle inspection task through spatial accurate modeling, importance scoring based on surface risk, reinforcement learning optimization of inspection path and dynamic adjustment of detection linkage, is significantly different from the existing inspection mode which relies on fixed route and has no real-time response, and effectively improves the inspection efficiency, defect discovery rate and power equipment maintenance level.

[0064] Further, the S1 comprises the following steps:

[0065] Autonomous scanning of the power grid area by several unmanned aerial vehicles equipped with laser radars and cameras to obtain laser point cloud data and two-dimensional image sequences of the power grid area;

[0066] 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] Multi-view image fusion processing of the two-dimensional image sequences to obtain a texture feature image of the power grid area, and mapping the texture feature image into the preliminary three-dimensional point cloud model to obtain a three-dimensional model of the power grid area.

[0068] In some embodiments, the region to be detected is first subjected to high-precision digital reconstruction. Specifically, several multi-rotor unmanned aerial vehicles (UAVs) equipped with laser radar (LiDAR) and high-resolution visible light cameras are deployed to perform multi-pass autonomous flight scanning operations on the target region. The flight tasks are executed according to pre-set geographical zoning and route planning, ensuring reasonable overlapping observation areas among different UAVs to improve modeling density and image fusion quality. Each UAV continuously acquires laser point cloud data and multi-angle two-dimensional image sequences within the target region during flight. The point cloud data is used to describe the three-dimensional spatial structure of the object surface in the target space, while the image sequences are used for subsequent texture reconstruction processing. The acquired raw point cloud data has a certain degree of noise and redundant point distribution due to factors such as laser reflectivity, flight attitude disturbance, and environmental occlusion, so filtering processing is required, including outlier rejection, ground separation, boundary sparsification, and normal vector reconstruction operations. Subsequently, multi-view point cloud registration algorithms (such as ICP or NDT) are used to coordinate and finely align the point clouds from different UAVs, ultimately generating a preliminary three-dimensional point cloud model covering the complete power grid region structure. This model can clearly depict the spatial layout relationship of devices such as transformers, switch cabinets, disconnectors, bus corridors, and transmission tower poles. At the same time, after performing image enhancement, distortion removal, and feature point extraction processing on the two-dimensional image sequences collected by the UAVs, multi-view image fusion is achieved using methods such as structure light beam adjustment and multi-view stereo reconstruction (MVS), reconstructing a texture feature image with high spatial resolution. This image can present surface detail features such as transformer body surface rust, insulator damage, and connection part corrosion in local areas. After obtaining the texture feature image, a texture mapping algorithm (such as a triangular mesh or voxel projection method) is used to accurately map the texture image onto the surface of the preliminary three-dimensional point cloud model, so that each spatial device structure surface has visual feature information, ultimately generating a complete three-dimensional textured power grid region model. This model not only has accurate geometric restoration capability but also retains surface texture details of power equipment, providing precise spatial foundation data for subsequent inspection object importance scoring, path planning, and surface defect recognition. This modeling method differs from existing two-dimensional perception strategies that rely on a single UAV and mainly use planar images, achieving high-precision three-dimensional spatial perception based on multi-source data fusion, with significant precision improvement and information integrity advantages.

[0069] Further, the S2 comprises the following steps:

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

[0071] According to the last inspection result data, a defect remaining data of the last inspection cycle is obtained by extracting un-repaired defect positions and un-inspected positions.

[0072] The initial hotspot data is spatially fused with the defect remaining data to obtain a hotspot fusion data set;

[0073] The hotspot fusion data set is input 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, the inspection object in the three-dimensional model of the power grid region is processed for prediction scoring to obtain a three-dimensional power grid model marked with the importance score of the inspection object.

[0075] In some embodiments, first, based on the information such as the location of power equipment defect occurrence, equipment type, defect category and occurrence time recorded in the historical defect database, a defect event feature vector jointly represented by spatial position and event frequency is constructed. On this basis, a density clustering algorithm is introduced to perform unsupervised spatial clustering processing on the above feature vector. Specifically, the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) method is selected, by setting the minimum sample number and spatial radius parameters, to identify the high-density aggregation area of defect events in geographical space, and to remove sparse abnormal points, finally to extract the hot spot area set of regular defect high incidence in the power grid area, which is recorded as the initial hot spot data. The initial hot spot data only reflects the fault distribution statistical characteristics in the historical period, and cannot reflect the potential risk remaining factors in the current period. In order to further enhance the perception ability of the model to the current task situation, the un-repaired defect locations and un-inspected points recorded in the last inspection process are extracted. The un-repaired defects are obtained by comparing the device ID and geographical position with the historical fault label and current processing state, and the un-inspected points are extracted by comparing the task scheduling log and planned path. The above two types of data are encoded into a spatial point set and given a high-risk label, and merged into the initial hot spot data to form a hot spot fusion data set containing structural historical aggregation characteristics and task residual factors. Then, the data samples for training the scoring model are constructed. Based on each point in the fusion data set as a sample unit, its spatial coordinates, device type, historical defect intensity indicators (such as annual average defect occurrence frequency), near defect point number density and state label are extracted to form a multi-dimensional feature vector, and a corresponding supervised label is constructed to represent the inspection priority score of the point. After dividing the above sample data into training set and validation set, input them into the pre-set convolutional neural network (CNN) model for training. The input layer of the CNN model accepts the two-dimensional or three-dimensional tensor structure of the inspection area encoding map, the convolution layer extracts spatial structure features and local risk distribution patterns using multiple convolution kernels, the pooling layer performs feature compression, and the fully connected layer generates importance scores of corresponding inspection points at the output end. The loss function adopts mean square error (MSE) or weighted regression error function for convergence optimization, and the score deviation and ordering consistency of the validation set results are verified after training to ensure that the model has good generalization ability and actual scoring reliability. Finally, the trained scoring model is deployed to the system task scheduling module to perform scoring inference operation on each inspection object in the power grid area three-dimensional model. The system generates an input feature vector based on the spatial position and equipment operation data of the inspection object, and inputs it into the CNN model to output the corresponding importance score value, which is automatically labeled to the corresponding object node of the power grid three-dimensional model. The scoring result is expressed in the form of heat value or risk weight, and serves as an important reference for subsequent path optimization and scheduling strategy.Compared with the traditional strategy based on geographic location or artificial experience weight scoring, the method realizes risk identification based on efficient allocation of inspection resources by fusing historical clustering distribution and current residual data and introducing a deep learning model for risk space feature modeling.

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

[0077] Based on the hotspot fusion data set, the position coordinates, the number of historical defect occurrences, the latest inspection defect detection result and the equipment type information of each inspection object are extracted to construct a training data set containing the spatial features, the historical defect features and the type features of the inspection objects;

[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] The training set is input into the convolutional neural network, and feature extraction is performed through the convolutional layer to obtain multi-scale spatial feature maps;

[0080] The multi-scale spatial feature maps are subjected to pooling dimension reduction processing to obtain compressed feature vectors;

[0081] The compressed feature vectors are input into a fully connected layer, regression training is performed, and the importance score of the corresponding inspection object is output;

[0082] The trained convolutional neural network is subjected to prediction accuracy verification through the validation set.

[0083] In some embodiments, in the data preprocessing stage, the basic attribute information of each inspection object is extracted from the fusion dataset one by one, including the position coordinates (X, Y, Z) in the three-dimensional space, the number of defect occurrences in the historical record of the device, the output results of the defect detection model in the last inspection process (including confidence, defect type), and the type code of the device (such as transformer, circuit breaker, disconnector, line support, etc.), which together constitute the 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 vector is standardized to make each dimension feature in a unified numerical range. Then, according to the importance label corresponding to each inspection object in the fusion dataset (generated by expert experience or inspection history scoring system), the whole data is divided into training set and validation set, usually in the ratio of 8:2 or 7:3, to ensure that the training set fully covers the feature distribution of each type of sample. The label value is expressed in the form of continuous numerical value, which is suitable for regression task modeling in neural network. In terms of model design, the convolutional neural network constructed is composed of multiple layers of stacked convolution modules, where the input layer receives the feature map of the inspection object after vector rearrangement and mapping to tensor structure, the convolution layer extracts the interaction features between local spatial adjacency relationship and multi-dimensional attributes through different size convolution kernels, forms multi-scale spatial feature map, and captures the potential high-order patterns of the inspection object in the space, time and device attribute dimensions layer by layer. After feature extraction, the pooling layer is connected for dimension reduction to compress redundant information and enhance the translation invariance of the model, and the maximum pooling (MaxPooling) or average pooling (Average Pooling) operation is often used. The pooling output is further input into the fully connected layer as deep semantic features, and the feature fusion and regression modeling are performed, and finally the importance score value of each inspection object is output. In the training process, the loss function selects the mean squared error (Mean Squared Error, MSE) as the evaluation index to measure the deviation between the network prediction score and the true label, and the Adam or RMSProp algorithm is used as the optimizer to improve the gradient convergence efficiency. In order to prevent overfitting, the Dropout regularization strategy is introduced, and the early stopping mechanism is used to monitor the change of validation set error to dynamically terminate the training process. After the model training is completed, the prediction accuracy and ranking consistency are evaluated on the validation set, and the verification indicators include mean absolute error, determination coefficient (R 2) and Spearman rank correlation coefficient, 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 input features of devices to be evaluated in any power grid area in the inference stage and output corresponding importance score results. The score results will be an important reference factor for the input of subsequent path planning algorithms, realizing the priority ranking of unmanned aerial vehicle inspection based on risk level. Compared with traditional scoring methods that rely on static rule setting or manual experience grading, the present scheme has self-learning ability, can adapt to complex power equipment structure and defect evolution characteristics, and significantly improves the intelligent level of inspection strategy and the dynamic response ability of the scoring mechanism

[0084] Further, the S3 comprises the following steps:

[0085] Based on the three-dimensional model of the power grid area, the spatial position and the corresponding importance score of each inspection object are extracted, and the initial position, power state, flight speed and inspection object defect confirmation state of each unmanned aerial vehicle are combined to construct an inspection task state space;

[0086] Based on the inspection task state space, the action space is defined as the action set of each unmanned aerial vehicle selecting a target inspection object in the next step, and the action set includes action options for approaching the inspection object at different headings and different distances;

[0087] The inspection task state space and the action space are input into a reinforcement learning model, and the next inspection target point of each unmanned aerial vehicle is output through action decision, obtaining a preliminary unmanned aerial vehicle inspection path sequence;

[0088] Based on the preliminary unmanned aerial vehicle inspection path sequence, the inspection path is optimized through a reward function, and the reward value of the reward function is calculated according to the inspection time, the number of discovered defect targets, the missed detection rate and the defect image clarity;

[0089] According to the strategy optimization result, the reinforcement learning model parameters are updated and a real-time unmanned aerial vehicle inspection route planning strategy is generated.

[0090] In some embodiments, firstly, a state space in the reinforcement learning task is constructed. The state space is based on the current flight state of each UAV and the global task distribution, including: the spatial position, power reserve, flight speed, target accessibility index of each UAV at the current time step, and the three-dimensional coordinates, importance score and defect confirmation state (such as whether it has been confirmed, whether it is a left-over high-risk target, etc.) of all inspection objects. The above state parameters jointly form a time-series state vector as the input of the policy network. In terms of action space design, a set of optional actions is defined for each UAV, and the set elements are the inspection target points that the UAV can choose to go to next and the corresponding flight strategy combination. The action set not only considers whether the target point position and the remaining power support round trip, but also combines the current heading angle offset and the flight cost to form a discrete action encoding structure, which typically includes candidate action options such as "go to the point with the maximum importance score", "go to the nearest un-inspected target", "perform hovering and waiting", etc. To enhance the model convergence efficiency, the size of the action space is limited by the Top-N pruning strategy during the training phase to reduce the search complexity. In the policy learning process, the above state space and action space are used as inputs, and the policy gradient method based on the deep reinforcement learning framework is used for policy optimization. The Actor-Critic architecture is selected, the policy network (Actor) accepts the state vector and outputs the action probability distribution, and the inspection action at the current time step is sampled based on the distribution; the value function network (Critic) evaluates the value of the state-action pair, which is used to guide the policy gradient update. The evaluation of the inspection path sequence depends on the reward function, which calculates the cumulative reward value after each round of path execution. The reward function designed in this scheme fully considers the actual needs of multi-objective optimization of the inspection task, and constructs the comprehensive reward value in the form of weighted combination, including the following sub-indexes: first, the inspection task completion efficiency, which is defined as a function of negative inspection time; second, the defect discovery ability, which is constructed as a positive reward based on the number of defect targets discovered per unit time; third, the task coverage completeness, which is measured as the complement of the omission rate (1-omission rate); fourth, the image quality evaluation index, which scores the images collected near the defect area based on the image sharpness score (such as based on structural similarity SSIM or resolution mean), encouraging high-quality image acquisition behavior. The final reward function is represented as a weighted linear combination of the above sub-indexes, and is processed by regularization to stabilize the training process. The reinforcement learning model uses the experience replay mechanism and the target network structure during training to improve the diversity of training samples and the stability of convergence. After training, the optimal policy model is deployed to the task scheduling system, which receives state input and outputs the next target action of each UAV in real time during actual operation, dynamically generating a preliminary inspection path sequence.During the running process, the state space is updated in real time after each current action is executed and the defect detection result is obtained, and the strategy network is called for the next round of action decision according to the new inspection image and the target state feedback information, to form a dynamic path optimization closed loop in time steps. Compared with the traditional inspection system based on fixed path table, the reinforcement learning path planning method has significant adaptability and task sensitivity, can dynamically optimize resource allocation according to the equipment risk level, task execution state and defect feedback information, maximize the inspection efficiency and defect discovery quality, and has strong algorithm innovation and application effectiveness.

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

[0092]

[0093] Wherein, R is the reward value; T is the cumulative time for the unmanned aerial vehicle to complete the inspection task; T ref is the preset reference time; D is the number of defect targets successfully found in the inspection process; L is the missed detection rate of the inspection task; Q is the defect image definition score collected by approaching the defect target; Q max is the maximum reference value of the definition score; a1, a2 and a4 are reward item coefficients; a3 is a penalty item coefficient.

[0094] It should be noted that the reward function first takes the proportional relationship between the actual time consumed by the unmanned aerial vehicle to complete the inspection task and the reference time set by the system as the basis to measure the timeliness of task execution. When the actual time is lower than the reference value, a positive reward will be obtained, otherwise a negative effect will be generated, thereby guiding the policy network to preferentially select the inspection sequence with shorter path and faster execution. Secondly, an index representing the "number of discovered defect targets" is introduced in the reward function to positively encourage the model to focus on areas with high defect risk or high-score equipment, thereby improving the overall defect discovery rate. In addition, considering the possibility of omission due to insufficient path coverage in the inspection task, the reward function further introduces the complement of the omission rate index, i.e., reflecting the integrity of the inspection coverage in the form of task completion rate. This index can guide the model to balance time efficiency and risk priority while maximizing global coverage to prevent the model from ignoring other targets due to excessive focus on high-score areas. Furthermore, to ensure that the detected images are interpretable and have defect recognition value, the reward function also includes an image clarity score item to reflect the image quality level collected by the unmanned aerial vehicle when approaching the equipment surface. The image quality index can be based on image clarity evaluation algorithms such as structural similarity index (SSIM) or gradient intensity value to reflect whether the image details are clear and whether it has defect feature recognition ability. This incentive mechanism encourages the strategy to adopt appropriate flight angles and heights when approaching high-risk targets to obtain high-quality images. Overall, the reward function is a weighted combination of multiple sub-indexes, and the coefficients of each item are adjusted according to the task objectives to achieve a balance between policy optimization under multiple target tasks. Through the gradual feedback and optimization process of the above reward function, the reinforcement learning model can adaptively learn and generate an inspection path strategy with high execution efficiency, high defect coverage rate, and image acquisition quality guarantee. Compared with traditional strategies that only optimize based on the shortest path or the number of defect points, this reward mechanism realizes more comprehensive judgment ability and practical engineering applicability of intelligent scheduling improvement.

[0095] Further, the S4 comprises the following steps:

[0096] S41, based on the real-time generated inspection route planning strategy, extracting the inspection action sequence of each unmanned aerial vehicle, generating an inspection task instruction set containing inspection target position, action instruction and execution sequence;

[0097] S42, based on the inspection task instruction set, scheduling and controlling each unmanned aerial vehicle to fly autonomously according to the action sequence, and synchronously collecting flight state data and inspection image stream data at each time step;

[0098] S43, input the inspection image stream data into a defect detection model for real-time defect detection to obtain a defect detection result of the current time step, and update an 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 to generate an inspection route planning strategy, and update an 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 result and the inspection target position information, perform defect position labeling in a three-dimensional model of the power grid area, and store the defect image, the equipment number and the detection time in association to generate an inspection data record.

[0101] In some embodiments, first, the inspection action sequence required to be performed by each UAV in the current task round is extracted from the path planning strategy output by the reinforcement learning model. The action sequence includes a plurality of action instructions with spatial target identification, each instruction explicitly corresponding to the spatial position code of the target device, the required heading angle of the UAV flight, the preset flight speed, the hovering time when shooting, and the sequence logic of task execution, thereby forming a structured inspection task instruction set, ensuring the timing and trajectory controllability of task execution. The system then calls the flight control interface to schedule and control each UAV according to the inspection task instruction set. When executing each action instruction, the UAV needs to complete the close flight and attitude adjustment of the specified target device under the support of accurate positioning and trajectory planning, and collect multi-view images within the preset hovering time. In order to ensure the data closed-loop capability, the UAV synchronously collects its flight state 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, and the image data is real-time returned to the edge processing unit or cloud reasoning server in the form of video frames or continuous static images. The collected image stream data is immediately sent to the deployed defect detection model in the edge computing unit for identification processing. The model based on deep convolutional neural network architecture extracts features and makes classification judgments for surface defects of power equipment (such as rust, cracks, insulation damage, etc.), and outputs the defect detection results of each target device at the current time step, including defect type, position bounding box, and confidence score. The detection results are not only used as the basis for fault judgment, but also feedback for dynamic updating of the inspection task state space. Specifically, when the system identifies that a device has clear defect signs and has not been confirmed and processed, the device node state will be updated from "to be inspected" to "marked", and a high priority label for inspection will be assigned; if the current target is confirmed to be normal, the corresponding node state will be changed to "completed" to reduce subsequent repeated scheduling. The updated task state space is re-input into the reinforcement learning path planning module to trigger a new round of strategy generation process. The reinforcement learning model dynamically adjusts the path planning logic in combination with the current flight state, uncompleted task point distribution, and image recognition results, and generates a new inspection task instruction set, realizing the adaptive scheduling of the whole process. The system determines whether there are still inspection tasks to be completed according to the instruction set: if the instruction set is not empty, the system continues to execute the inspection by returning to the scheduling control module; if the instruction set is empty, the system considers that the current task round has been completely covered, and enters the post-processing phase. After the task is completed, the system labels the target position of the discovered defect in the power grid area three-dimensional model according to the final defect detection results and device spatial information, and generates a defect distribution map with spatial positioning information. Each defect information is bound with its corresponding image data, device number, detection time, detection model output confidence, etc. parameters, and is written into the inspection data record system.The above record can be used for subsequent fault repair scheduling, device operation and maintenance analysis, and risk prediction model training.

[0102] Further, the defect detection model is deployed on a cloud server in communication connection with the unmanned aerial vehicle.

[0103] It should be noted that, in order to balance the model accuracy and real-time requirements of the power equipment surface defect identification task, the defect detection model is deployed in a cloud server with high-performance computing capability, and data transmission and inference result interaction are performed through a low-latency communication link established with the unmanned aerial vehicle system. The model uses a deep convolutional neural network architecture, has multi-scale texture perception and location boundary recognition capability, and is dedicated to identifying common defect features such as cracks, rust, damage, insulator contamination, and loose wires that may appear on the surface of power grid equipment. During the execution of the inspection task, the unmanned aerial vehicle pre-processes the image stream data through the edge module (such as size redefinition, frame extraction, noise filtering, etc.), and uploads the processed image information in the form of compressed data packets to the cloud detection platform in real time through 5G or LTE wireless communication channels. The cloud server immediately invokes the deployed defect detection model for online inference after receiving the image data, and outputs structured detection results including defect type, location coordinates, confidence score, and suspicious target region image mask. The detection process is completed in the cloud in batch inference or parallel GPU acceleration mode, ensuring low latency and high throughput in large-scale image stream processing.

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

[0105] Based on the historical image data of power equipment inspection in the power grid, original image samples are extracted, and the image samples are labeled to obtain a training dataset and a validation dataset containing defect target category and location box labels;

[0106] Based on the training dataset, perform size normalization, color normalization, and data enhancement processing on the image samples to obtain pre-processed training image data;

[0107] A Transformer network model is constructed, and the network parameters are initialized;

[0108] The pre-processed training image data is input into the Transformer network model, and the model parameters are optimized by joint training of weighted cross-entropy loss and boundary 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 construction of the defect detection model.

[0110] Specifically, first, based on the power equipment inspection history image database, raw image samples covering various types of equipment surface states are screened out, including typical defect instances such as transformer shell cracks, circuit breaker contact surface rust, power line holder loosening, insulator contamination spots, and lightning arrester damage. For the extracted raw images, a semi-automatic labeling platform is used to manually label each image in fine granularity, including the category label and corresponding location information of the defect target. The defect area is marked in the form of a bounding box, and finally a training dataset and a validation dataset containing image data, category label and location box annotation are constructed. In the data preprocessing stage, in order to improve the training stability and generalization ability of the model, standardization processing is performed on the image samples in the training dataset. In the size normalization process, all images are scaled and padded according to a fixed input resolution (such as 512x512 or 640x640 pixels) to adapt to the fixed input size requirement of the self-attention mechanism in the Transformer model. The color normalization operation unifies the mean and standard deviation of each channel to eliminate the data bias caused by light changes in different shooting environments. To further enhance the robustness of the model to defect features under different scales, angles, and lighting conditions, multiple data augmentation strategies are performed, including random horizontal flipping, random affine transformation, color jittering, and random cropping, to expand the training sample space. The main structure of the model uses a target detection model based on the Transformer encoder-decoder architecture. First, the network parameters are initialized, including the weight matrix of the self-attention layer and the feedforward neural network module in the encoder, the query vector in the decoder, and the output head layer weight, and the stability of the initial training is improved through moderate parameter normalization. The preprocessed training images are input into the Transformer network. The image is first extracted by a shallow convolutional network to obtain a preliminary feature map, and then sent to the encoder for multi-layer multi-head self-attention feature modeling to capture long-range dependency and local texture features. The decoder predicts the defect instances in the image based on a set of learnable query vectors and outputs the class probability distribution and corresponding bounding box coordinate parameters of each target. During training, the optimization goal is to minimize the classification loss and positioning loss. The classification loss uses a weighted cross-entropy function, and to address the problem of uneven distribution of defect classes, the rare defect classes are given higher weights to improve the detection accuracy of small sample classes. The positioning loss uses a bounding box regression loss such as GIoU or DIoU loss to optimize the position matching degree between the predicted box and the real box. The overall loss function is the weighted sum of the classification loss and the regression loss, and the model parameters are iteratively optimized through the backpropagation algorithm. After training, the model performance is evaluated using an independent validation dataset, and the evaluation indicators include precision (Precision), recall (Recall), mean average precision (mAP), and positioning error (IoU mean).According to the performance feedback results of the verification set, the model hyperparameters are fine-tuned, including learning rate adjustment, regularization coefficient optimization and Transformer encoder layer selection, to further improve the detection stability and positioning accuracy of the final model in the actual inspection environment. Through the above complete process, the Transformer deep learning model suitable for power equipment surface defect detection task is finally completed, which has self-adaptive recognition ability to various complex surface defect features, and significantly outperforms traditional convolutional network in long-distance feature modeling and multi-target detection accuracy.

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

[0112] A power grid model module is used to construct a three-dimensional model of the power grid area.

[0113] An inspection marking module is used to mark the importance score of the inspection object in the three-dimensional model of the power grid area based on the historical defect data of the power equipment in the power grid area through a neural network model.

[0114] An intelligent planning module is used to construct a UAV inspection route planning strategy based on the importance score of the inspection object and the three-dimensional model of the power grid area through a reinforcement learning model.

[0115] A defect detection module is used to schedule the UAV inspection according to the UAV inspection route planning strategy, and simultaneously acquire the inspection image stream for power equipment defect detection.

[0116] The above embodiments only describe the preferred embodiments of the application, and do not limit the scope of the application. Without departing from the design spirit of the application, various modifications and improvements to the technical solutions of the application made by ordinary engineering technicians in the art shall fall within the protection scope determined by the claims of the application.

Claims

1. A deep learning-based power equipment defect detection method, characterized in that, The method comprises the following steps: S1, constructing a three-dimensional model of a power grid area; S2, based on historical defect data of power equipment in the power grid area, marking the importance score of the inspection object in the three-dimensional model of the power grid area through a neural network model; S3, based on the three-dimensional model of the power grid area, constructing a UAV inspection route planning strategy through a reinforcement learning model; S4, scheduling the UAV according to the UAV inspection route planning strategy, and synchronously acquiring an inspection image stream for power equipment defect detection; The S2 comprises the following steps: Based on the historical defect data of the power equipment in the power grid area, a clustering algorithm is used for spatial hotspot analysis to obtain initial hotspot data of a high-incidence area of conventional defects in the power grid area; According to the last inspection result data, the un-repaired defect positions and un-inspected positions are extracted to obtain defect remaining data of the last inspection period; The initial hotspot data and the defect remaining data are spatially fused to obtain a hotspot fusion data set; The hotspot fusion data set is input 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, the inspection object in the three-dimensional model of the power grid area is predicted and scored to obtain a three-dimensional power grid model marked with the importance score of the inspection object.

2. The power equipment defect detection method based on deep learning according to claim 1, characterized in that, The S1 comprises the following steps: A plurality of UAVs carrying laser radars and cameras are used to autonomously scan the power grid area to obtain laser point cloud data and a two-dimensional image sequence of the power grid area; Based on the laser point cloud data, filtering and registration processing is performed to obtain a preliminary three-dimensional point cloud model of the power grid area; Multi-view image fusion processing is performed on the two-dimensional image sequence to obtain a texture feature image of the power grid area, and the texture feature image is mapped into the preliminary three-dimensional point cloud model to obtain a three-dimensional model of the power grid area. 3.The power equipment defect detection method based on deep learning of claim 1, wherein, The convolutional neural network is constructed by the following steps: Based on the hotspot fusion data set, the position coordinates, historical defect occurrence times, last inspection defect detection results and equipment type information of each inspection object are extracted to construct a training data set containing spatial features, historical defect features and type features of the inspection object; 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; The training set is input into the convolutional neural network, and feature extraction is performed through the convolutional layer to obtain a multi-scale spatial feature map; The multi-scale spatial feature map is subjected to pooling and dimension reduction processing to obtain a compressed feature vector; The compressed feature vector is input into a fully connected layer, regression training is performed, and the importance score of the corresponding inspection object is output; The trained convolutional neural network is subjected to prediction accuracy verification through the validation set. 4.The power equipment defect detection method based on deep learning of claim 1, wherein, The S3 comprises the following steps: Based on the three-dimensional model of the power grid area, the spatial position and corresponding importance score of each inspection object are extracted, and the initial position, power state, flight speed and inspection object defect confirmation state of each UAV are combined to construct an inspection task state space; Based on the inspection task state space, define the action space as the action set of each UAV selecting a target inspection object in the next step, the action set including action options for approaching the inspection object at different headings and different distances; Input the inspection task state space and the action space into the reinforcement learning model, output the next inspection target point of each UAV through action decision, and obtain a preliminary UAV inspection path sequence; Based on the preliminary UAV inspection path sequence, perform policy optimization on the inspection path through a reward function, the reward value of the reward function being calculated according to the inspection time, the number of defect targets found, the omission rate and the defect image clarity; According to the policy optimization result, update the reinforcement learning model parameters and generate a real-time UAV inspection route planning strategy.

5. The power equipment defect detection method based on deep learning according to claim 4, characterized in that, The reward function is as follows: ； wherein, is a reward value; is a cumulative time for the UAV to complete the inspection task; is a preset reference time; is a number of defect targets successfully found in the inspection process; is a missed detection rate of the inspection task; is a definition score of a defect image collected close to the defect target; is a maximum reference value of the definition score; , and is a reward term coefficient; is a penalty term coefficient.

6. The power equipment defect detection method based on deep learning according to claim 4, characterized in that, The S4 includes the following steps: S41, based on the real-time UAV inspection route planning strategy, extract the inspection action sequence of each UAV, and generate an inspection task instruction set containing inspection target positions, action instructions and execution sequences; S42, based on the inspection task instruction set, control each UAV to fly autonomously according to the action sequence, and synchronously collect flight state data and inspection image stream data at each time step; S43, input the inspection image stream data into a defect detection model for real-time defect detection, obtain the defect detection result of the current time step, and update the inspection task state space based on the defect detection result; S44, based on the updated inspection task state space, input the reinforcement learning model again to 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 result and the inspection target position information, mark the defect position in the three-dimensional model of the power grid area, and store the defect image, equipment number and detection time in association to generate an inspection data record.

7. The power equipment defect detection method based on deep learning according to claim 6, characterized in that, The defect detection model is deployed on a cloud server in communication connection with the UAV.

8. The power equipment defect detection method based on deep learning according to claim 7, characterized in that, The defect detection model is constructed by the following steps: Based on the historical image data of power equipment inspection in the power grid, extract original image samples, and label the image samples to obtain a training data set and a verification data set containing defect target category and position box labels; Based on the training data set, perform size normalization, color normalization and data enhancement processing on the image samples to obtain preprocessed training image data; Construct a Transformer network model and initialize network parameters; Input the preprocessed training image data into the Transformer network model, train through weighted cross-entropy loss and boundary box regression loss, and optimize model parameters; Evaluate the detection accuracy and positioning accuracy of the Transformer network model based on the verification data set, and adjust the model hyperparameters according to the evaluation result to complete the construction of the defect detection model. 9.A deep learning-based power equipment defect detection system, applied to the deep learning-based power equipment defect detection method of any one of claims 1-8, characterized in that, It includes: A power grid model module for constructing a three-dimensional model of a power grid area; The inspection marking module is configured to mark the importance score of the inspection object on the three-dimensional model of the power grid area based on the historical defect data of the power equipment in the power grid area through a neural network model. The intelligent planning module is configured to construct a UAV inspection route planning strategy based on the importance score of the inspection object and the three-dimensional model of the power grid area through a reinforcement learning model. The defect detection module is configured to perform inspection scheduling on the UAV according to the UAV inspection route planning strategy, and synchronously acquire an inspection image stream for power equipment defect detection.

Images