A typhoon post-disaster unmanned aerial vehicle intelligent patrol decision method based on image recognition and semantic analysis

Abstract

This invention discloses an intelligent UAV patrol decision-making method based on image recognition and semantic analysis after a typhoon disaster. The method includes: acquiring multi-source post-typhoon data and preprocessing it to form multi-modal post-disaster basic data; performing semantic analysis on disaster reporting and repair text data to extract typhoon characteristic information, affected equipment information, and prone-to-fault information to generate prior disaster information; fusing and analyzing the prior disaster information with equipment operation data to determine the suspected damage status of each piece of equipment or section to be inspected and key inspection targets; generating UAV patrol routes by combining equipment ledger data, meteorological data, and UAV operation constraints; controlling the UAV to collect target images and performing image enhancement and image recognition to obtain equipment damage identification results; generating post-disaster patrol decision results based on the identification results and transmitting them back to the repair command platform. This invention can improve the accuracy of post-disaster patrol target selection, the robustness of image recognition, and the efficiency of repair resource allocation.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) inspection technology, specifically relating to an intelligent UAV inspection decision-making method based on image recognition and semantic analysis after a typhoon disaster. Background Technology

[0002] Following extreme weather disasters such as typhoons, power grid facilities typically suffer widespread physical damage, including collapsed towers, damaged insulators, broken conductors, detached hardware, and blocked access routes. Rapid and accurate post-disaster equipment damage assessment and inspection are crucial for guiding the allocation of emergency repair resources, minimizing power outage time, and restoring the lifeline of power supply. In recent years, drones, due to their high mobility, wide field of view, and ability to access areas inaccessible by humans, have been increasingly used in post-typhoon power grid inspections. Meanwhile, the handling of past typhoon disasters has accumulated a wealth of heterogeneous data from multiple sources, including meteorological data, equipment operation data, disaster reports, repair logs, work orders, customer service repair requests, and drone inspection images. Effectively utilizing this data to improve the accuracy of post-disaster inspection decisions and equipment damage identification has become a significant technical challenge in power grid post-disaster recovery.

[0003] Existing post-disaster inspection and damage assessment methods mainly include the following categories: One category is the fixed-route drone inspection method based on human experience, where maintenance personnel pre-set the inspection route based on experience, and after the drone completes the image or video collection, the human staff plays back the video frame by frame or browses the photos in the background to find the fault point; Another category is to introduce image recognition or edge computing modules into the drone inspection process to perform real-time or near-real-time recognition on the collected inspection images to help discover anomalies in transmission lines, towers, insulators or channel environments; and yet another category is to archive and manage historical emergency repair records, fault records or disaster reporting information, and use keyword retrieval, statistical analysis or simple data mining methods to assist in post-disaster assessment.

[0004] For example, existing technology CN118865179A discloses a method for drone-based disaster reconnaissance after a typhoon. This method acquires raw data of power transmission lines after a typhoon, performs data mining and feature extraction on this data to obtain characteristic data of the transmission lines, and generates a disaster reconnaissance and inspection strategy accordingly. Simultaneously, based on the monitoring data and positioning information of each drone, it triggers the drones to conduct post-typhoon disaster inspections according to the strategy. It also utilizes a lightweight edge computing module based on embedded AI for detection and recognition, mounted on the drones, to perform real-time edge computing, and receives the visualized inspection data from the drones via a remote integrated image / data transmission wireless link. This type of technology can improve the automation level of drone-based disaster reconnaissance after a typhoon to a certain extent and reduce the workload of relying entirely on manual inspections.

[0005] However, existing technologies still have shortcomings in dealing with complex power grid recovery scenarios after typhoons. Current UAV disaster reconnaissance methods mostly rely on data mining and feature extraction of raw data from transmission lines to generate inspection strategies. They typically focus more on the status characteristics of the line side and lack in-depth semantic analysis of the large amount of unstructured information generated during past typhoon disasters, such as disaster reports, repair logs, work orders, and customer service repair requests. This makes it difficult to fully extract the empirical correlation patterns between typhoon characteristics, affected equipment, and common faults, and results in insufficient utilization of historical disaster response experience.

[0006] Secondly, existing inspection strategies typically focus on generating disaster relief tasks based on transmission line characteristic data. They do not adequately consider the comprehensive correlation between meteorological forecasts and measured data, equipment malfunctions, equipment ledger topology relationships, and historical damage patterns. This makes it difficult to accurately distinguish the damage risks and inspection urgency of different equipment or sections in situations where there is widespread damage after a typhoon and the number and endurance of drones are limited. This can easily lead to inaccurate allocation of inspection resources.

[0007] Furthermore, the post-typhoon disaster environment is typically characterized by complex interferences such as heavy rainfall, dense fog, low light levels, waterlogging and glare, fallen trees obstructing the view, and drone vibration. While existing image recognition methods based on embedded AI or lightweight edge computing can improve on-site processing efficiency, they may still be affected by factors such as image quality degradation, target occlusion, and unclear edges of small defects in the complex post-disaster context. The accuracy and robustness of identifying fine-grained damaged parts such as insulator breakage, tower cracks, conductor strand breakage, and vibration damper displacement still need to be improved.

[0008] Furthermore, while existing technologies can utilize drone monitoring data, location information, and wireless links to transmit patrol data, they typically lack comprehensive consideration of factors such as post-disaster micro-meteorological conditions, wind resistance, energy consumption, no-fly zones, multi-drone coordination, and return-to-base safety in drone flight path planning. In situations where post-typhoon conditions involve complex wind fields, blocked roads, and intensive patrol missions, failure to adequately consider these factors in flight path planning may result in drones being unable to prioritize coverage of high-risk areas, or problems such as insufficient endurance, redundant patrols, and flight path conflicts.

[0009] Meanwhile, existing post-disaster drone-based disaster reconnaissance methods primarily focus on acquiring patrol data and identifying faults during or after the disaster. They lack sufficient consideration for evaluating the recovery effectiveness after repairs, analyzing the return on investment in repairs, and providing feedback for subsequent disaster prevention planning. Data on substantial repair resource input, power restoration effectiveness, critical section repair status, and power restoration response efficiency have not been effectively quantified and evaluated. This makes it difficult to promptly identify issues such as redundant material deployment, delayed personnel response, or insufficient output in different repair areas, and also hinders the effective support for subsequent windproofing reinforcement, spare parts reserves, and the layout of material storage facilities.

[0010] Therefore, existing technologies still have technical problems such as insufficient depth of multi-source data fusion after disasters, insufficient utilization of unstructured emergency repair experience, insufficient precision in target selection for UAV inspections, insufficient robustness of image recognition in complex post-disaster backgrounds, insufficient consideration of meteorological and energy consumption constraints in flight path planning, and lack of closed-loop quantitative evaluation of post-disaster recovery effects. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for intelligent inspection and decision-making by unmanned aerial vehicles (UAVs) after typhoon disasters based on image recognition and semantic analysis.

[0012] The objective of this invention can be achieved through the following technical solutions: This invention provides a method for intelligent post-typhoon drone patrol decision-making based on image recognition and semantic analysis, comprising the following steps: S1. Acquire multi-source post-typhoon data and preprocess the multi-source post-typhoon data to obtain multi-modal post-typhoon basic data; wherein, the multi-source post-typhoon data includes meteorological forecast and measured data, equipment operation data, equipment ledger data, disaster reporting and emergency repair text data, and post-typhoon image data collected by UAVs. S2. Perform semantic analysis on the disaster reporting and emergency repair text data, extract typhoon characteristic information, affected equipment information and prone fault information related to typhoon disaster, and generate disaster prior information based on the typhoon characteristic information, the affected equipment information and the prone fault information; S3. The disaster prior information is integrated and analyzed with the equipment operation data to determine the suspected damage status of each piece of equipment or section to be inspected, and the key inspection targets are determined based on the suspected damage status and the equipment ledger data. S4. Based on the key inspection targets, the equipment ledger data, the meteorological forecast and measured data, and the drone operation constraints, generate drone inspection tasks and corresponding drone inspection routes. S5. Control the UAV to collect target images according to the UAV patrol route, and perform image enhancement and image recognition on the target images to obtain equipment damage identification results; wherein, the equipment damage identification results include the location of the damaged equipment, the damaged part and the damage type; S6. Generate post-disaster inspection decision results based on the key inspection targets, the UAV inspection route, and the equipment damage identification results, and send the post-disaster inspection decision results back to the emergency repair command platform to guide the allocation of post-disaster emergency repair resources.

[0013] Furthermore, the preprocessing of the post-disaster multi-source data to obtain post-disaster multimodal basic data specifically includes: The meteorological forecast and measured data are processed into grids and time-corrected to extract meteorological features corresponding to each piece of equipment or section to be inspected; wherein, the meteorological features include precipitation, temperature, humidity, typhoon intensity, typhoon path, wind speed and wind direction; Physical boundary screening is performed on the equipment operation data to remove abnormal data that exceeds the physical range of equipment operation, and spatiotemporal completion is performed on missing data caused by communication interruption or collection anomaly to obtain preprocessed equipment operation data. For short-term missing data in the time dimension, interpolation is performed using equipment operation data from adjacent time points to complete the missing data. Isolated anomaly detection is performed on the equipment operation data after spatiotemporal completion to remove isolated noise data generated by false alarms from post-disaster sensors, resulting in denoised equipment operation data; The equipment ledger data is subjected to unified equipment identification, standardized spatial location, and power grid topology verification to obtain preprocessed equipment ledger data; The disaster reporting and emergency repair text data are cleaned, the equipment name is standardized, the location name is standardized, and the timestamp is corrected. The standardized equipment name and location name are then matched with the preprocessed equipment ledger data for fuzzy matching and key-value mapping to obtain text intelligence data corresponding to the ledger equipment. Video frame extraction is performed on the post-disaster image data collected by the UAV to obtain candidate image frames, and the sharpness of the candidate image frames is evaluated based on the Laplace variance. Candidate image frames with a sharpness evaluation value lower than a preset sharpness threshold are identified as invalid image frames and removed, while valid post-disaster image data are retained; Using the preprocessed equipment ledger data as the correlation benchmark, the meteorological features, the denoised equipment operation data, the text intelligence data, and the effective post-disaster image data are correlated according to equipment identification, spatial location, and minute-level time information to obtain the post-disaster multimodal basic data.

[0014] Furthermore, the semantic analysis of the disaster reporting and emergency repair text data, extracting typhoon characteristic information, affected equipment information, and prone-to-fault information related to the typhoon disaster, specifically includes: The disaster reporting and emergency repair text data are segmented into sentences, invalid characters are removed, equipment names are standardized, fault description words are standardized, and place names are standardized to obtain a standardized text sequence. The normalized text sequence is input into a pre-trained language representation model, and dynamic word vector representations incorporating contextual semantics are obtained through self-attention computation. The language representation model is a BERT model built on a Transformer encoder. The BERT model includes a word embedding layer, a position embedding layer, a fragment embedding layer, and a multi-layer Transformer encoding layer, which are used to generate dynamic word vector representations based on the contextual relationships of each word in the normalized text sequence. The dynamic word vector representation is input into the bidirectional sequence feature extraction model to extract the forward context features and backward context features of the normalized text sequence, respectively. The forward context features and backward context features are then concatenated to obtain a comprehensive semantic feature vector. The bidirectional sequence feature extraction model is a BiLSTM model built on a bidirectional long short-term memory network. The BiLSTM model includes a forward long short-term memory network and a backward long short-term memory network. The forward long short-term memory network is used to extract the semantic dependencies from front to back in the normalized text sequence, and the backward long short-term memory network is used to extract the semantic dependencies from back to front in the normalized text sequence. The comprehensive semantic feature vector is input into the conditional random field decoding layer, and the normalized text sequence is globally labeled according to the label transition relationship to obtain a text label sequence; wherein, the text label sequence includes typhoon feature labels, equipment labels, component labels, fault labels, handling labels, and location labels; The conditional random field decoding layer aims to maximize the conditional probability of the text label sequence, and the conditional probability is given by the formula: in, Indicates the input text sequence Output text label sequence under certain conditions The conditional probability; This represents the normalized text sequence; This represents the sequence of text tags to be decoded; Represents any sequence of candidate text labels; Indicates the first The comprehensive semantic feature vector corresponding to each word element; Represents the text label sequence The Middle The tags corresponding to each word element; Represents the candidate text label sequence The Middle The tags corresponding to each word element; Representation and Label The corresponding feature weight parameters; Representation and Label The corresponding feature weight parameters; Indicates by label Transfer to label Tag transfer parameters; Indicates by label Transfer to label Tag transfer parameters; Based on the text tag sequence, extract typhoon characteristic entities, equipment entities, component entities, and fault entities from the normalized text sequence; Based on the semantic relationships between the typhoon characteristic entity, the equipment entity, the component entity, and the fault entity in the same disaster record or emergency repair record, a structured triple is generated; wherein, the structured triple includes typhoon characteristics, affected equipment, and prone faults; Based on the structured triples, the typhoon characteristic information, the affected equipment information, and the prone fault information are obtained.

[0015] Furthermore, disaster prior information is generated based on the typhoon characteristic information, the affected equipment information, and the prone-to-fault information, specifically including: Based on the typhoon characteristic information, the affected equipment information, and the prone fault information, a structured triplet set consisting of typhoon characteristics, affected equipment, and prone faults is constructed. The affected equipment in the structured triplet set is matched with the equipment ledger data by equipment identifier and spatial location to determine the equipment or section to be inspected corresponding to each affected equipment. Based on the correspondence between typhoon characteristics, equipment or sections to be inspected, and common faults, the conditional co-occurrence frequency between different entities is statistically analyzed, and a prior knowledge matrix is ​​generated. The elements in the prior knowledge matrix are represented as follows: in, Represents the first in the prior knowledge matrix The fault type corresponding to the equipment or section to be inspected. The prior probability of historical vulnerability; Represents a set of fault types; In the set of structured triples, the first... The equipment or section to be inspected and the type of fault. The number of times conditional co-occurrence occurs; This represents the Laplace smoothing coefficient, and ; Represents the set of fault types The number of fault types; Based on the prior knowledge matrix, determine the historical vulnerability prior probability and the corresponding common fault type of each equipment or section to be inspected; The prior probability of historical vulnerability, the type of common failure, and the typhoon characteristics corresponding to the type of common failure are combined to generate the prior disaster information; wherein, the prior disaster information includes the equipment or section to be inspected, the prior probability of historical vulnerability, the typhoon characteristics, and the type of common failure.

[0016] Furthermore, S3 specifically includes: The historical vulnerability prior probability, typhoon characteristics, and prone failure types corresponding to each equipment or section to be inspected are obtained from the disaster prior information. The electrical abnormality signals corresponding to each piece of equipment or section to be inspected are extracted from the equipment operation data to form an electrical abnormality signal set; wherein, the electrical abnormality signals include line impedance abnormality, protection action signal, power outage alarm signal, telemetry over-limit signal, current abnormality signal, voltage abnormality signal and communication interruption signal; Based on the power grid topology connection relationship between the equipment or sections to be inspected, a Bayesian network is constructed to represent the correlation between disaster prior information, electrical anomaly signals and equipment physical damage events; Based on the Bayesian network, the prior probability of historical vulnerability is jointly inferred with the set of electrical anomaly signals to determine the posterior probability of damage for each piece of equipment or section to be inspected. The posterior probability of damage is given by the following formula: in, Indicates the electrical abnormality signal set Conditions, No. The posterior probability of physical damage to a piece of equipment or section to be inspected; Indicates the first An event in which physical damage occurs to the equipment or section to be inspected; This represents the set of electrical anomaly signals extracted from the operating data of the equipment; Indicates the first The conditional probability of the electrical abnormality signal set occurring when a piece of equipment or section to be inspected suffers physical damage; This refers to the set of equipment or sections to be inspected that are participating in the joint inference; Indicates the first The prior historical vulnerability probability corresponding to each piece of equipment or section to be inspected is expressed as: in, Represents the first in the prior knowledge matrix The fault type corresponding to the equipment or section to be inspected. The prior probability of historical vulnerability; Indicates the first A set of common fault types corresponding to each piece of equipment or section to be inspected; Indicates the first The fault type corresponding to the equipment or section to be inspected. Fault association weights; Based on the posterior damage probability and in conjunction with the common fault types in the prior disaster information, the suspected damage status of each piece of equipment or section to be inspected is determined; wherein, the suspected damage status includes the suspected damage probability, the suspected damage level, and the suspected fault type; Based on the equipment ledger data, the importance of each piece of equipment or section to be inspected is determined. Then, based on the posterior probability of damage and the equipment importance, a high-risk urgency index for drone inspection of each piece of equipment or section to be inspected is determined. The formula for the high-risk urgency index for drone inspection is: in, Indicates the first The urgency index of high-risk drone inspection of individual equipment or sections to be inspected; Indicates the first The importance of each piece of equipment or section to be inspected; This represents the weighting coefficient corresponding to the posterior probability of damage; This represents the weighting coefficient corresponding to the importance of the device; Equipment or sections to be inspected that reach a preset urgency threshold for UAV inspection high-risk urgency index are identified as candidate high-risk targets. These candidate high-risk targets are then merged and deduplicated based on their spatial location, power grid topology connection relationship, and suspected fault type to obtain key inspection targets. The key inspection targets include target equipment identification, target spatial location, suspected damage level, suspected fault type, and UAV inspection high-risk urgency index.

[0017] Furthermore, the step of generating drone inspection tasks and corresponding drone inspection routes based on the key inspection targets, the equipment ledger data, the meteorological forecast and measured data, and drone operation constraints specifically includes: Based on the target equipment identification and target spatial location in the key inspection targets, the corresponding equipment type, line, section, power grid topology connection relationship and equipment installation height are matched in the equipment ledger data to generate a set of high-risk inspection nodes; Based on the target spatial location, equipment installation height, suspected fault type, and UAV patrol urgency index of each high-risk patrol node in the set of high-risk patrol nodes, the corresponding three-dimensional patrol waypoints are determined; wherein, the three-dimensional patrol waypoints include waypoint latitude and longitude, flight altitude, hovering shooting position, and shooting attitude; Based on the meteorological forecasts and measured data, obtain the wind speed, wind direction, precipitation and visibility of the flight segment between any two three-dimensional patrol points, and determine the wind resistance impact of the corresponding flight segment in combination with the UAV flight direction; Based on the wind resistance effect, flight distance, and flight time, the segment energy consumption between any two three-dimensional survey waypoints is determined. The segment energy consumption is calculated using the following formula: in, Indicates that the drone is from the first The three-dimensional survey waypoint flew to the first Energy consumption of each three-dimensional waypoint patrol route; This indicates the airspeed energy consumption coefficient of the drone; Indicates the drag penalty coefficient; This indicates the speed of the drone relative to the air. This indicates the wind speed corresponding to the stated flight segment; Indicates the angle between the drone's flight direction and the wind direction; Based on the set of high-risk inspection nodes, the energy consumption of the flight segment, and the constraints of UAV operations, a multi-UAV collaborative path optimization model is constructed; wherein, the multi-UAV collaborative path optimization model aims to prioritize covering high-risk inspection nodes with a high urgency index for UAV inspections and reduce the energy consumption of the flight segment. The multi-objective ant colony algorithm is used to solve the multi-UAV cooperative path optimization model. In the nth iteration, the 1st Only one ant from the first The 3D survey waypoint was transferred to the first... The state transition probability of a 3D sighting waypoint is given by the formula: in, Indicates the first In the nth iteration Only one ant from the first The 3D survey waypoint was transferred to the first... The state transition probability of each three-dimensional waypoint; Indicates the first In the nth iteration The three-dimensional survey waypoint and the first Pheromon concentration between three-dimensional waypoints; Indicates the first The three-dimensional survey waypoint to the first Heuristic factors for three-dimensional waypoints; Indicates the first The urgency index of high-risk drone patrols corresponding to each three-dimensional patrol waypoint; Indicates the first The set of three-dimensional waypoints that an ant is allowed to access in its current state; This indicates the influencing factors corresponding to pheromone concentration; This represents the impact factor corresponding to the heuristic factor; This indicates the influencing factors corresponding to the high-risk urgency index of drone patrols; The heuristic factor is determined based on the flight distance and segment energy consumption between three-dimensional waypoints, and is expressed as: in, Indicates the first The three-dimensional survey waypoint to the first Flight distance of each three-dimensional waypoint; This indicates the penalty weight corresponding to the energy consumption of a flight segment; Under the condition of satisfying the UAV operation constraints, the access order of the three-dimensional inspection waypoints corresponding to each UAV is iteratively optimized to generate UAV inspection tasks; wherein, the UAV inspection task includes the UAV number, the target to be inspected, the access order of the three-dimensional inspection waypoints and the corresponding shooting task. According to the UAV patrol mission, the take-off point, three-dimensional patrol waypoint, hovering shooting point and return point of each UAV are connected in sequence to generate the UAV patrol route; wherein, the UAV patrol route is a three-dimensional flight waypoint sequence including latitude and longitude, altitude, access order, shooting attitude and execution UAV number.

[0018] Furthermore, the drone operation constraints include: The power constraint limits the sum of the energy consumption of each UAV during its patrol mission (segment energy consumption, hovering energy consumption, and return energy consumption) to no more than the available power of that UAV. The power constraint is expressed as follows: in, Indicates the first The drone patrol route corresponding to each drone; This refers to a segment formed by two adjacent three-dimensional waypoints in the UAV's patrol route; Indicates the first The hovering energy consumption required for a drone to perform hovering photography mission; Indicates the first The energy consumption required for a drone to return to its landing point after completing an inspection; Indicates the first The available battery power of the drone; This represents the power safety margin coefficient; Flight distance constraints and flight time constraints are used to limit the total flight distance and total flight time of each UAV corresponding to the UAV patrol route to not exceed the maximum allowable flight distance and maximum allowable flight time of the corresponding UAV. Meteorological safety constraints are used to limit the UAV from performing patrol missions within meteorological areas that meet safe flight conditions; wherein, the meteorological safety constraints include wind speed not exceeding the maximum permissible wind speed, precipitation not exceeding the maximum permissible precipitation, and visibility not being lower than the minimum permissible visibility; Airspace safety constraints are used to restrict the UAV's patrol route from avoiding no-fly zones, temporary control zones, and obstacle zones, and to maintain a preset safe distance between the UAV and poles, power lines, buildings, and other UAVs; Flight altitude constraints are used to limit the flight altitude of each three-dimensional inspection waypoint to within the allowable altitude range based on the equipment installation height, the type of equipment to be inspected, and the shooting requirements in the equipment ledger data. Multi-drone collaborative constraints are used to limit the same high-risk inspection node to a maximum of one drone in the same inspection round, and to restrict different drones from entering the same spatial conflict area within the same time period; Return-to-home constraints are used to ensure that each UAV can return to the preset landing point or emergency landing point from the last three-dimensional waypoint after completing its corresponding UAV inspection mission. The task coverage constraint is used to prioritize drone coverage of high-risk inspection nodes where the drone inspection urgency index reaches a preset urgency threshold, and to determine drone inspection tasks to be executed in descending order of the drone inspection urgency index when the number of drones or battery power is insufficient.

[0019] Furthermore, S5 specifically includes: The drone is controlled to reach the three-dimensional inspection waypoint corresponding to the key inspection target according to the drone inspection route, and the image is acquired based on the hovering shooting position and shooting posture corresponding to the three-dimensional inspection waypoint to obtain the target image; The target image is associated with the corresponding acquisition time, drone location, shooting posture, key patrol targets, and equipment ledger data to obtain the target image to be identified; The image to be identified is subjected to an image degradation state judgment; wherein, the image degradation state includes low light, heavy fog, heavy rain, water flooding reflection, motion blur and occlusion interference; When the target image to be identified is in a state of image degradation, anti-degradation image enhancement processing is performed on the target image to be identified at the edge computing node of the UAV or the emergency repair command platform to obtain an enhanced target image; wherein, the anti-degradation image enhancement processing includes global restoration processing based on generative image enhancement and local contrast enhancement processing based on adaptive histogram. The enhanced target image is input into the damaged part recognition model to detect the power grid equipment, equipment components and damaged areas in the enhanced target image, and to obtain the equipment detection box, the damaged part detection box, the damage type and the recognition confidence. Based on the device detection frame, the damaged part detection frame, the recognition confidence level, the drone position, the shooting posture, and the device ledger data, the location of the damaged device, the damaged part, and the damage type are determined, and the device damage recognition result is generated.

[0020] Furthermore, the anti-deterioration image enhancement processing and the damaged area identification model specifically include: The anti-deterioration image enhancement processing includes conditional generative adversarial network enhancement processing and contrast-limited adaptive histogram equalization processing. The conditional generative adversarial network includes a generator network and a discriminator network. The generator network is used to map a target image to be identified that has deteriorated into a clear image. The discriminator network is used to distinguish between a real clear image and an enhanced image generated by the generator network. The enhanced image is subjected to contrast-limited adaptive histogram equalization processing. By limiting and redistributing the gray-level distribution of local image regions, the local texture of insulator skirts, hardware bolts, tower connections and conductor edges is enhanced to obtain the enhanced target image. The damaged area identification model is a single-stage target detection model that incorporates a collaborative attention mechanism. The single-stage target detection model includes a backbone feature extraction network, a collaborative attention module, and a detection head. The enhanced target image is input into the backbone feature extraction network to extract deep convolutional feature maps; The collaborative attention module includes a channel attention module and a spatial attention module; The channel attention module generates channel attention weights based on the global average pooling and global max pooling results of the deep convolutional feature maps. Based on the channel attention weights, the deep convolutional feature map is reconstructed to obtain a channel-reconstructed feature map. The spatial attention module generates spatial attention weights based on the average pooling and max pooling results of the channel reconstructed feature maps; Based on the spatial attention weights, the channel reconstruction feature map is spatially reconstructed to obtain an attention-enhanced feature map. The attention-enhanced feature map is input into the detection head to identify the damaged area in the enhanced target image, and the device detection box, the damaged part detection box, the damage type, and the recognition confidence are output; wherein, the damage type includes insulator damage, insulator string breakage, tower deformation, tower crack, conductor strand breakage, missing hardware, and vibration damper displacement.

[0021] Furthermore, S6 specifically includes: Based on the key inspection targets, the UAV inspection route, and the equipment damage identification results, a post-disaster inspection decision result is generated; wherein, the post-disaster inspection decision result includes damaged equipment identification, damaged equipment location, damaged part, damage type, identification confidence level, inspection priority, emergency repair priority, and emergency repair resource requirements; The post-disaster inspection decision results are transmitted back to the emergency repair command platform, which then allocates emergency repair personnel, vehicles, emergency supplies, power generation equipment, and drone verification tasks according to the emergency repair priority and the emergency repair resource requirements. After the emergency repair task is completed, decision-making units are divided according to the emergency repair responsibility area or emergency repair grid area, and the emergency repair input indicators and emergency repair output indicators corresponding to each decision-making unit are collected. The emergency repair input indicators include the time spent on manual labor, the number of emergency repair vehicles, the number of emergency lighting devices, the number of power generation devices, the amount of windproof and reinforcement materials invested, and the amount of spare parts invested. The emergency repair output indicators include the restoration of power load, the number of key transmission sections repaired, the number of equipment repaired, the power restoration response speed, and the number of faults eliminated. Based on the emergency repair input indicators and the emergency repair output indicators, an input indicator matrix and an output indicator matrix are constructed, and the BCC model in data envelopment analysis is used to calculate the comprehensive post-disaster recovery efficiency value of each decision-making unit. For the For each decision-making unit, the comprehensive efficiency value of post-disaster recovery is calculated according to the following formula: in, Indicates the first The overall efficiency value of post-disaster recovery for each decision-making unit; Represents a non-Archimedean infinitesimal; This indicates the quantity of emergency repair inputs; This indicates the quantity of emergency repair output indicators; Indicates the number of decision-making units participating in the evaluation; Indicates the first The decision-making unit corresponds to the first... The indicator value of emergency repair investment; Indicates the first The decision-making unit corresponds to the first... The indicator value of emergency repair investment; Indicates the first The decision-making unit corresponds to the first... The indicator value of each emergency repair output indicator; Indicates the first The decision-making unit corresponds to the first... The indicator value of each emergency repair output indicator; Indicates the first The intensity coefficient corresponding to each decision-making unit when constructing the evaluation benchmark; Indicates the first Redundant slack variables corresponding to each emergency repair input indicator; Indicates the first The output insufficiency slack variable corresponding to each emergency repair output indicator; Based on the comprehensive post-disaster recovery efficiency value, the input redundancy slack variable, and the output deficiency slack variable, the post-disaster recovery effectiveness evaluation results of each decision-making unit are determined; wherein, the post-disaster recovery effectiveness evaluation results include the effective recovery area, the input redundancy area, the output deficiency area, and the response lag area; Based on the post-disaster recovery effectiveness evaluation results, disaster prevention and mitigation feedback information is generated; wherein, the disaster prevention and mitigation feedback information includes information on adjustments to windproofing and reinforcement needs, information on adjustments to emergency repair resource allocation, information on adjustments to spare parts reserves, and information on optimization of material reserve warehouse locations; The disaster prevention and mitigation feedback information is fed back into the pre-disaster prevention and control plan and subsequent typhoon disaster inspection decision-making, and is used to update the disaster prior information, the constraints of drone operations, the demand for emergency repair resources, and the rules for determining the key inspection targets.

[0022] Compared with the prior art, the present invention has the following advantages: (1) Although existing technologies can use data mining to assist in generating disaster relief and inspection strategies, they mostly remain at the level of keyword retrieval or simple archiving for unstructured data such as emergency repair logs, disaster reports, work order records, and customer service repair reports generated during past typhoon disasters. It is difficult to extract deep-seated correlation patterns between typhoon characteristics, affected equipment, and common faults. This invention extracts typhoon characteristic information, affected equipment information, and common fault information by performing BERT dynamic semantic representation, BiLSTM contextual feature extraction, and conditional random field global label decoding on disaster reports and emergency repair text data. Furthermore, it forms structured triples, transforming historical emergency repair experience, which was originally difficult to directly participate in calculations, into structured knowledge that can be used for post-disaster judgment, thereby improving the ability of historical disaster experience to support current post-disaster inspection decisions.

[0023] (2) Existing technologies typically rely on the current characteristics of power transmission lines or human experience when determining the targets for UAV inspections, making it difficult to fully utilize historical damage patterns. This results in certain historically vulnerable equipment or frequently faulty sections not being prioritized for identification during current post-disaster inspections. This invention constructs a priori knowledge matrix based on typhoon characteristic information, affected equipment information, and fault-prone information. Furthermore, it generates disaster priori information based on the conditional co-occurrence relationships between different equipment or sections and different fault types. This allows each piece of equipment or section to be inspected to obtain the corresponding historical vulnerability prior probability and fault-prone type. Therefore, the system can not only focus on equipment that has already shown abnormalities but also identify potential high-risk targets with historical vulnerabilities in advance, improving the foresight of post-disaster inspection target selection.

[0024] (3) Existing technologies lack sufficient joint judgment between real-time equipment operation anomalies and historical disaster experience during the generation of post-disaster inspection strategies. This can easily lead to misjudgments based solely on current alarms or delays in inspections based solely on historical experience. This invention integrates disaster prior information with equipment operation data, performs Bayesian joint inference based on electrical anomaly signal sets and historical vulnerability prior probabilities, determines the posterior damage probability of each piece of equipment or section to be inspected, and combines the equipment importance in the equipment ledger data to form a high-risk urgency index for UAV inspection. Thus, even in the case of widespread disaster and complex alarm information after a disaster, it can simultaneously consider the probability of damage and the degree of impact on the power grid, allowing limited UAV resources to prioritize covering equipment or sections that are more likely to be damaged and have a greater impact.

[0025] (4) Existing UAV patrol routes typically focus more on path distance or fixed patrol tasks, neglecting practical operational factors such as wind speed, wind direction, rainfall, visibility, no-fly zones, UAV battery power, and multi-UAV collaboration after typhoons. This can easily lead to a mismatch between the route planning results and the actual flight conditions at the disaster site. This invention generates UAV patrol tasks and routes based on key patrol targets, equipment ledger data, meteorological forecasts and measured data, and UAV operational constraints. It also introduces wind resistance penalties and flight energy consumption factors into the segment energy consumption calculation, and uses a multi-target ant colony algorithm to optimize the order of visiting three-dimensional patrol waypoints. As a result, it can improve the coverage efficiency of high-risk nodes and reduce redundant patrols, ineffective detours, and mission interruptions due to insufficient battery power, while ensuring flight safety and feasible endurance.

[0026] (5) While existing drones equipped with lightweight AI modules can achieve on-site recognition, the post-typhoon environment is often accompanied by low light, heavy fog, heavy rainfall, flooding and glare, tree obstruction, and motion blur, resulting in severe image quality degradation and unclear edges of small defects, thus affecting the recognition results. This invention assesses the image degradation state of the target image before image recognition, and performs global restoration processing based on generative image enhancement and local contrast enhancement processing based on adaptive histograms when degradation is present, making details such as insulator skirts, hardware bolts, tower connections, and conductor edges clearer. This reduces the interference of complex post-disaster backgrounds on the image recognition model and improves the usability of image data after extreme weather.

[0027] (6) Existing basic image recognition algorithms are prone to misjudging background interference as equipment defects when faced with obstructions from fallen tree branches, reflections from accumulated water, mudslide backgrounds, and minor cracks in equipment, or may miss localized damage such as broken conductor strands or displacement of vibration dampers. This invention introduces a single-stage target detection model with an enhanced target image input and a collaborative attention mechanism. Channel attention highlights feature responses related to the material, contour, and components of power grid equipment, while spatial attention focuses on suspected damaged areas in the image. The model then outputs the location of the damaged equipment, the damaged part, the damage type, and the recognition confidence level. As a result, the accuracy of fine-grained defect localization and classification can be improved in complex post-disaster backgrounds, making the inspection results more suitable for direct use in emergency repair dispatching and on-site verification.

[0028] (7) Existing technologies focus primarily on post-disaster inspections and fault identification, neglecting the recovery effects, resource input efficiency, and feedback from subsequent disaster prevention planning. This leads to difficulties in quantifying and accumulating post-disaster experience, and subsequent typhoon prevention efforts may still suffer from problems such as redundant materials, delayed personnel response, or insufficient reserves in key areas. This invention, after the completion of a repair task, divides decision-making units according to the repair responsibility area or repair grid area, collects repair input and output indicators, and uses the BCC model in Data Envelopment Analysis to calculate the comprehensive post-disaster recovery efficiency value, thereby generating disaster prevention and mitigation feedback information. Thus, it can evaluate the relative efficiency of repair input and recovery output in different areas, identify areas with redundant input, insufficient output, and delayed response, and feed the evaluation results back to pre-disaster prevention planning, spare parts reserves, and subsequent typhoon disaster inspection decisions, achieving a closed-loop improvement between post-disaster inspections, repair assessments, and subsequent disaster prevention planning. Attached Figure Description

[0029] Figure 1 This is a flowchart of the intelligent inspection and decision-making method for typhoon disaster relief using unmanned aerial vehicles (UAVs) according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the generation of prior disaster information and the determination of key inspection targets according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the unmanned aerial vehicle (UAV) patrol route generation, image recognition, and decision feedback in an embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0031] like Figure 1 , Figure 2 , Figure 3 As shown, this embodiment provides a post-typhoon disaster drone intelligent patrol decision-making method based on image recognition and semantic analysis, including the following steps: S1. Obtain multi-source data after the typhoon disaster and preprocess the multi-source data to obtain multi-modal basic data after the disaster. In one specific implementation, post-disaster multi-source data includes meteorological forecasts and measured data, equipment operation data, equipment ledger data, disaster reporting and emergency repair text data, and post-disaster image data collected by drones. Meteorological forecasts and measured data can come from meteorological grid platforms, typhoon path forecast systems, and on-site micro-meteorological monitoring devices; equipment operation data can come from power transmission and distribution automation systems, online monitoring devices, and fault alarm systems; equipment ledger data can come from power grid GIS ledger systems; disaster reporting and emergency repair text data can come from disaster reporting records, emergency repair logs, work order records, and customer service repair reports; and post-disaster image data collected by drones can be inspection video streams or image frames.

[0032] Meteorological forecasts and measured data are processed into grids and time-corrected to map meteorological characteristics such as precipitation, temperature, humidity, typhoon intensity, typhoon path, wind speed, and wind direction to each piece of equipment or section to be inspected. This approach is necessary because wind and rain conditions vary significantly across different regions after a typhoon, and using only administrative region-level meteorological data would not accurately reflect the true disaster situation near specific lines, towers, or sections.

[0033] Physical boundary screening is performed on equipment operating data to remove obvious anomalies that exceed the equipment's permissible operating range, such as data related to voltage, current, temperature, or communication status that do not conform to the equipment's physical constraints. For short-term missing data caused by communication interruptions or data acquisition anomalies, interpolation between adjacent time points can be used to complete the data. The completion formula is as follows: in, Indicates the time to be completed The equipment operation data; Indicates time Known equipment operating data; Indicates time Known equipment operating data; and These represent the time points to be completed. The adjacent valid sampling times before and after.

[0034] For missing data in local grids in the spatial dimension, inverse distance weighted imputation can be performed based on the known data of adjacent grids. The imputation formula is as follows: in, Indicates time Lower target grid The completed data value; Indicates time Lower reference grid The known data values; Represents the target mesh The surrounding reference mesh set; Represents the target mesh With reference grid Spatial distance between them; This represents the distance attenuation coefficient. The closer the reference grid is, the greater its influence on the target grid, which better reflects the continuous spatial distribution of meteorological and equipment operating conditions, thereby improving the reliability of the missing data completion results.

[0035] After completing the spatiotemporal completion, isolated anomaly detection is performed on the equipment operation data to identify isolated noise data caused by post-disaster sensor false alarms, communication jitter, or instantaneous sampling anomalies. Specifically, the isolated forest algorithm can be used to score the anomalies of multi-dimensional equipment operation characteristics, and data with anomaly scores exceeding a preset threshold are removed as noise alarms. This process can reduce the interference of a large amount of post-disaster false alarm data on subsequent assessments of suspected damage status.

[0036] The equipment ledger data is standardized in terms of equipment identification, spatial location, and power grid topology. For example, the codes of the same tower, line, switch, or distribution transformer in different systems are unified into the same equipment identifier, and the latitude and longitude, elevation, line, section, and upstream and downstream connection relationships are verified to be consistent, so as to avoid the problem of the same name for different objects or the same object with different names when linking multiple source data in the future.

[0037] Text cleaning, equipment name standardization, location name standardization, and timestamp correction are performed on disaster reporting and emergency repair text data. For abbreviations, typos, non-standard place names, or colloquial equipment names appearing in manual reports, fuzzy matching is used to match them with equipment ledger data, establishing a key-value mapping relationship between equipment identifiers, spatial locations, and text records, thus obtaining text intelligence data corresponding to the equipment in the ledger. Through this processing, disaster descriptions in the text can be pinpointed to specific equipment or specific areas, providing reliable objects for subsequent semantic analysis.

[0038] Video frames are extracted from the post-disaster image data collected by drones to obtain candidate image frames. The sharpness of the candidate image frames is then evaluated based on the Laplace variance. The sharpness evaluation formula is as follows: in, Represents candidate image frames Clarity rating; Indicates candidate image frames; This represents the edge response result obtained after processing the candidate image frame using the Laplacian operator. This indicates variance calculation.

[0039] Image quality is assessed by leveraging the stronger edge response changes in sharp images and the weaker edge response changes in blurry images. When the sharpness evaluation value is below a preset sharpness threshold, the corresponding candidate image frame is identified as invalid and discarded; when the sharpness evaluation value reaches the preset sharpness threshold, it is retained as valid post-disaster image data. This reduces the number of invalid images caused by drone shake, heavy rain, and low visibility entering the subsequent recognition process, improving the stability of image recognition results.

[0040] Using preprocessed equipment ledger data as the correlation benchmark, meteorological characteristics, denoised equipment operation data, text intelligence data, and effective post-disaster image data are correlated according to equipment identification, spatial location, and minute-level time information to obtain post-disaster multimodal basic data. In this way, each piece of equipment or section to be inspected can be mapped to meteorological status, operational status, text disaster information, and image information within the same time window, providing a unified data foundation for subsequent semantic analysis, damage probability inference, flight path planning, and image recognition.

[0041] S2. Perform semantic analysis on disaster reporting and emergency repair text data to extract typhoon characteristic information, affected equipment information and prone failure information related to typhoon disasters, and generate disaster prior information based on typhoon characteristic information, affected equipment information and prone failure information; In one specific implementation, the disaster reporting and emergency repair text data includes disaster reporting records, repair logs, work order records, customer service repair requests, and on-site reporting information generated in the early stages of this typhoon disaster. Because this type of text often suffers from inconsistent equipment abbreviations, colloquial location descriptions, non-standard fault descriptions, and incomplete time records, the disaster reporting and emergency repair text data is first processed through sentence segmentation, invalid character removal, equipment name standardization, fault description term standardization, and place name standardization to obtain a standardized text sequence. This standardization process reduces recognition bias caused by the same equipment or the same fault being expressed in multiple ways in different texts.

[0042] Normalized text sequences are input into a pre-trained language representation model to generate dynamic word vector representations that incorporate contextual semantics. The language representation model employs a BERT model built on a Transformer encoder, which includes word embedding layers, positional embedding layers, fragment embedding layers, and multiple Transformer encoding layers. Word embedding layers represent the meaning of the words themselves, positional embedding layers preserve the order of words within a sentence, fragment embedding layers distinguish different text fragments, and multiple Transformer encoding layers capture the contextual dependencies between words. Compared to static word vectors, the BERT model can distinguish the meaning of the same word in different disaster contexts based on the context. For example, "broken wires" in different sentences may correspond to broken wire strands, communication interruption, or line outage.

[0043] In the BERT model, self-attention computation can be represented as: in, This represents the output of self-attention; This represents the query matrix obtained by mapping a normalized text sequence. This represents the key matrix obtained by mapping a normalized text sequence; This represents the value matrix obtained by mapping a normalized text sequence; Indicates transpose; Indicates the dimension of the key vector; This represents the normalized exponential function.

[0044] By calculating the relevance between the query matrix and the key matrix, attention weights are assigned to different lexical units, and then the value matrix is ​​used to generate a contextual semantic representation. With this setup, the model can automatically focus on word combinations in disaster texts that are more important for fault diagnosis, such as "typhoon path," "tower collapse," "insulator damage," and "conductor breakage."

[0045] The dynamic word vector representation is input into the bidirectional sequence feature extraction model to extract forward and backward context features of the normalized text sequence. These features are then concatenated to obtain a comprehensive semantic feature vector. The bidirectional sequence feature extraction model employs a BiLSTM model built upon a bidirectional long short-term memory network. The BiLSTM model includes a forward long short-term memory network and a backward long short-term memory network. The forward long short-term memory network extracts semantic dependencies from front to back, while the backward long short-term memory network extracts semantic dependencies from back to front. This setup is because the boundaries of fault entities in emergency repair texts often depend on the context for accurate identification. For example, in the sentence "a conductor strand broke near tower 27 of a certain line," the equipment object, spatial location, and fault type need to be considered in conjunction with the semantics of the entire sentence for accurate identification.

[0046] The comprehensive semantic feature vector is input into the Conditional Random Field (CRF) decoding layer. Based on the label transition relationships, the normalized text sequence is globally labeled to obtain a text label sequence. This text label sequence includes typhoon feature labels, equipment labels, component labels, fault labels, handling labels, and location labels. The CRF decoding layer can constrain the transition relationships between adjacent labels, avoiding label combinations that do not conform to entity labeling rules, and improving the accuracy of boundary identification for complex power entities and fault entities.

[0047] The Conditional Random Field (CRF) decoding layer aims to maximize the conditional probability of the text label sequence, which is expressed as: in, Indicates the input text sequence Output text label sequence under certain conditions The conditional probability; Represents a normalized text sequence; This represents the sequence of text tags to be decoded; Represents any sequence of candidate text labels; Indicates the first The comprehensive semantic feature vector corresponding to each word element; Represents a sequence of text labels The Middle The tags corresponding to each word element; Represents candidate text label sequence The Middle The tags corresponding to each word element; Representation and Label The corresponding feature weight parameters; Representation and Label The corresponding feature weight parameters; Indicates by label Transfer to label Tag transfer parameters; Indicates by label Transfer to label The label transfer parameter; by setting this parameter, the globally optimal label sequence can be selected across the entire sentence, rather than classifying individual words independently, thereby reducing entity boundary segmentation errors.

[0048] Based on the text label sequence, typhoon characteristic entities, equipment entities, component entities, and fault entities are extracted from the normalized text sequence. Typhoon characteristic entities may include information such as typhoon level, wind speed, wind direction, rainfall intensity, and path location; equipment entities may include objects such as lines, towers, transformers, switches, and substations; component entities may include components such as insulators, conductors, hardware, vibration dampers, and crossarms; and fault entities may include fault descriptions such as broken lines, collapsed towers, damage, cracks, displacement, missing parts, short circuits, and grounding.

[0049] Based on the semantic relationships between typhoon characteristic entities, equipment entities, component entities, and fault entities within the same disaster or emergency repair record, structured triples are generated. A structured triple includes typhoon characteristics, affected equipment, and common faults. For example, if an emergency repair record contains "strong winds," "a certain power line tower," and "broken conductor strands," a structured triple "strong winds—a certain power line tower—broken conductor strands" can be formed. In this way, historical disaster experiences in unstructured text are transformed into statistically statistically computable structured data.

[0050] Based on typhoon characteristic information, affected equipment information, and prone fault information, a structured triplet set is constructed, consisting of typhoon characteristics, affected equipment, and prone faults. The affected equipment in the structured triplet set is matched with equipment ledger data using equipment identifier and spatial location matching to determine the corresponding equipment or section to be inspected for each affected device. Through ledger matching, equipment descriptions in text can be accurately mapped to specific equipment nodes or line sections in the power grid GIS, preventing historical experience from remaining merely textual and unable to participate in subsequent inspection decisions.

[0051] Based on the correspondence between typhoon characteristics, equipment or sections to be inspected, and prone-to-fault conditions, the conditional co-occurrence frequency among different entities is statistically analyzed, and a prior knowledge matrix is ​​generated. The elements in the prior knowledge matrix are represented as follows: in, Represents the first in the prior knowledge matrix The fault type corresponding to the equipment or section to be inspected. The prior probability of historical vulnerability; Represents a set of fault types; In the set of structured triples, the first... The equipment or section to be inspected and the type of fault. The number of times conditional co-occurrence occurs; This represents the Laplace smoothing coefficient, and ; Represents the set of fault types The system calculates the number of fault types and, by determining the proportion of a particular fault type co-occurring with a specific device or section within the total number of fault co-occurrences, obtains the historical vulnerability prior probability for that fault type. This setup reflects the differences in vulnerability of different devices or sections during past typhoon disasters; for example, a coastal section is historically more prone to conductor strand breakage, while a mountainous section is more prone to tower tilting. The resulting prior knowledge matrix provides a historical experience basis for subsequent inferences about suspected damage conditions.

[0052] Based on the prior knowledge matrix, the historical vulnerability prior probabilities and corresponding prone failure types of each piece of equipment or section to be inspected are determined. The historical vulnerability prior probabilities, prone failure types, and corresponding typhoon characteristics are combined to generate disaster prior information. This disaster prior information includes the equipment or section to be inspected, the historical vulnerability prior probabilities, typhoon characteristics, and prone failure types. Through this process, textual experience from previous typhoon disasters can be utilized in advance before the current typhoon post-disaster inspection, ensuring that the determination of subsequent key inspection targets not only relies on current warnings but also takes into account historical vulnerability patterns.

[0053] S3. Integrate and analyze disaster prior information with equipment operation data to determine the suspected damage status of each piece of equipment or section to be inspected, and determine key inspection targets based on the suspected damage status and equipment ledger data. In one specific implementation, the historical vulnerability prior probability, typhoon characteristics, and prone fault types corresponding to each piece of equipment or section to be inspected are first read from the disaster prior information. The historical vulnerability prior probability reflects the empirical pattern of failures in similar equipment or sections in previous typhoon disasters, and the prone fault types are used to help determine the direction of the fault that the current anomaly is more likely to correspond to, such as broken conductor strands, tilted towers, damaged insulators, or missing hardware.

[0054] Electrical anomaly signals corresponding to each piece of equipment or section to be inspected are extracted from the equipment operation data to form an electrical anomaly signal set. These signals may include one or more of the following: abnormal line impedance, protection operation signals, power outage alarm signals, telemetry over-limit signals, abnormal current signals, abnormal voltage signals, and communication interruption signals. This process transforms real-time operational anomalies after a disaster into observational evidence that can be used for probabilistic inference, avoiding judgment delays caused by relying solely on historical experience.

[0055] A Bayesian network is constructed based on the power grid topology connections between the equipment or sections to be inspected. Nodes in the Bayesian network include nodes representing prior disaster information, nodes representing electrical anomaly signals, and nodes representing physical damage events. Edges between nodes represent the conditional dependencies between typhoon disasters, equipment vulnerability, operational anomalies, and physical damage. For example, if there are topological connections between upstream and downstream sections of the same line, a line break or ground fault in one section may trigger abnormal signals from adjacent switches, protection devices, or downstream equipment. Therefore, constructing an inference network based on the power grid topology connections can reduce misjudgments caused by isolated analysis of individual alarms.

[0056] Based on Bayesian networks, the prior probabilities of historical vulnerabilities are jointly inferred with a set of electrical anomaly signals to determine the posterior damage probability of each piece of equipment or section to be inspected. The posterior damage probability is expressed as: in, Indicates the electrical abnormality signal set Conditions, No. The posterior probability of physical damage to a piece of equipment or section to be inspected; Indicates the first An event in which physical damage occurs to the equipment or section to be inspected; This represents the set of electrical anomaly signals extracted from equipment operation data; Indicates the first The conditional probability of an electrical abnormality signal set occurring when a piece of equipment or section to be inspected suffers physical damage; This indicates the set of equipment or sections to be inspected that are participating in the joint inference; this setting can avoid false alarms caused by relying solely on current alarms, and can also avoid missed judgments caused by relying solely on historical experience, making the suspected damage status more consistent with the actual post-disaster operation.

[0057] in, Indicates the first The prior historical vulnerability probability corresponding to each piece of equipment or section to be inspected is expressed as: in, Represents the first in the prior knowledge matrix The fault type corresponding to the equipment or section to be inspected. The prior probability of historical vulnerability; Indicates the first A set of common fault types corresponding to each piece of equipment or section to be inspected; Indicates the first The fault type corresponding to the equipment or section to be inspected. Fault association weights; Based on the posterior probability of damage and combined with the common fault types in the prior disaster information, the suspected damage status of each piece of equipment or section to be inspected is determined. The suspected damage status includes the suspected damage probability, the suspected damage level, and the suspected fault type. The suspected damage level can be divided into high-risk, medium-risk, and low-risk according to the interval of the posterior probability of damage; the suspected fault type can be determined jointly based on the common fault types and the current electrical anomaly signals. For example, if the historically common fault is conductor strand breakage, and at the same time there are current line impedance anomalies and protection operation signals, the suspected fault type can be preferentially identified as a conductor-related fault.

[0058] The importance of each piece of equipment or section to be inspected is determined based on equipment inventory data. Equipment importance can be determined by considering factors such as voltage level, number of power users, whether it is associated with important users such as hospitals, emergency command centers, or transportation hubs, and whether it is located on critical interconnections or important transmission lines. For post-disaster recovery, targets with a high probability of damage but a small impact area should have different inspection priorities than targets with a high probability of damage and impact on important loads; therefore, it is necessary to further introduce equipment importance assessment.

[0059] Based on the posterior probability of damage and the importance of the equipment, a high-risk urgency index for drone inspection is determined for each piece of equipment or section to be inspected. The high-risk urgency index for drone inspection is expressed as: in, Indicates the first The urgency index of high-risk drone inspection of individual equipment or sections to be inspected; Indicates the first The importance of each piece of equipment or section to be inspected; This represents the weighting coefficient corresponding to the posterior probability of damage. This represents the weighting coefficient corresponding to the importance of the equipment; The probability of damage and the degree of impact on the power grid are combined into a unified inspection ranking index. The posterior probability of damage reflects whether the target may have already suffered physical damage, while the equipment importance reflects the impact on power restoration and critical load protection should the target be damaged. With this setting, given the limited number of drones, battery life, and post-disaster flight windows, priority can be given to drones inspecting equipment or sections that are more likely to be damaged and have a greater impact, thereby improving the efficiency of limited inspection resources.

[0060] Equipment or sections requiring inspection by drone that reach a preset urgency threshold are identified as candidate high-risk targets. Multiple candidate high-risk targets that are spatially adjacent, belong to the same line section, have the same suspected fault type, or are closely connected in the power grid topology are merged and deduplicated to obtain key inspection targets. Key inspection targets include target equipment identification, target spatial location, suspected damage level, suspected fault type, and drone inspection urgency index. Merging and deduplication reduces redundant drone flights and photography of adjacent targets, making subsequent flight path planning more compact and facilitating the emergency repair command platform to organize segment-based verification and repair.

[0061] S4. Based on key inspection targets, equipment ledger data, meteorological forecasts and measured data, and drone operation constraints, generate drone inspection tasks and corresponding drone inspection routes. In one specific implementation, based on the target equipment identification and spatial location among the key inspection targets, the corresponding equipment type, associated line, associated section, power grid topology connection relationship, and equipment installation height are matched against the equipment ledger data to generate a high-risk inspection node set. Each node in the high-risk inspection node set corresponds to a piece of equipment or line section that requires priority inspection by drones, and retains the corresponding suspected damage level, suspected fault type, and drone inspection urgency index. This setup allows flight path planning to consider both equipment damage risk and power grid operation impact, rather than solely relying on spatial distance.

[0062] Based on the target spatial location, equipment installation height, suspected fault type, and UAV inspection urgency index of each high-risk inspection node in the high-risk inspection node set, three-dimensional inspection waypoints are determined for each high-risk inspection node. The three-dimensional inspection waypoints include the waypoint's latitude and longitude, flight altitude, hovering shooting position, and shooting attitude. For example, for a target suspected of insulator damage, the hovering shooting position can be set in the lateral visible area of ​​the insulator string; for a target suspected of conductor strand breakage, the shooting attitude can be arranged along the direction of conductor extension. By combining the suspected fault type to set the three-dimensional inspection waypoints, the coverage of the damaged area in subsequent image acquisition can be improved, and invalid images generated by the UAV at the scene due to unsuitable angles can be reduced.

[0063] Based on meteorological forecasts and measured data, wind speed, wind direction, precipitation, and visibility are obtained for any two three-dimensional patrol waypoints, and the wind resistance impact of the corresponding flight segment is determined in conjunction with the UAV's flight direction. After a typhoon, wind fields vary significantly across different corridors within the same region. If flight path planning only considers geometric distance, problems such as short paths with severe headwinds, excessive energy consumption, or poor flight stability may occur. Therefore, micro-meteorological factors need to be incorporated into the flight segment cost before generating flight paths.

[0064] Based on the effects of wind resistance, flight distance, and flight time, the segment energy consumption between any two three-dimensional waypoints is determined, and the segment energy consumption is expressed as: in, Indicates that the drone is from the first The three-dimensional survey waypoint flew to the first Energy consumption of each three-dimensional waypoint patrol route; This indicates the airspeed energy consumption coefficient of the drone; Indicates the drag penalty coefficient; This indicates the speed of the drone relative to the air. Indicates the wind speed corresponding to the flight segment; Indicates the angle between the drone's flight direction and the wind direction; When a drone flies against the wind or is affected by strong crosswinds, the wind resistance penalty increases, leading to higher energy consumption for that flight segment. This setting allows route planning to proactively avoid segments with excessive energy consumption or high flight risks, improving mission feasibility under complex weather conditions after a disaster.

[0065] Based on the set of high-risk inspection nodes, flight segment energy consumption, and UAV operational constraints, a multi-UAV collaborative path optimization model is constructed. This model aims to prioritize coverage of high-risk inspection nodes with a high urgency index while reducing flight segment energy consumption, while also considering task allocation and flight path conflicts among UAVs. For post-typhoon inspection scenarios, the number of UAVs, battery life, and safe flight windows are all limited; therefore, path optimization cannot simply pursue the shortest distance but must strike a balance between prioritizing coverage of high-risk targets and controlling flight costs.

[0066] A multi-objective ant colony algorithm is used to solve the multi-UAV cooperative path optimization model. In the ttt-th iteration, the t-th iteration... Only one ant from the first The 3D survey waypoint was transferred to the first... The state transition probability of a 3D sighting waypoint is expressed as: in, Indicates the first In the nth iteration Only one ant from the first The 3D survey waypoint was transferred to the first... The state transition probability of each three-dimensional waypoint; Indicates the first In the nth iteration The three-dimensional survey waypoint and the first Pheromon concentration between three-dimensional waypoints; Indicates the first The three-dimensional survey waypoint to the first Heuristic factors for three-dimensional waypoints; Indicates the first The urgency index of high-risk drone patrols corresponding to each three-dimensional patrol waypoint; Indicates the first The set of three-dimensional waypoints that an ant is allowed to access in its current state; This indicates the influencing factors corresponding to pheromone concentration; This represents the impact factor corresponding to the heuristic factor; This indicates the influencing factors corresponding to the high-risk urgency index of drone patrols; In this state transition probability, pheromone concentration is used to retain search experience from better paths in historical iterations, heuristic factors are used to guide the UAV to select flight segments with better distance and energy consumption, and the UAV inspection urgency index is used to increase the probability of high-risk targets being prioritized for access. Compared with path selection based solely on distance, this setting enables UAVs to prioritize covering more likely and more important nodes under limited endurance conditions after a disaster, reducing the crowding out of inspection resources for high-risk targets by low-risk targets.

[0067] The heuristic factor is determined based on the flight distance between waypoints and the segment energy consumption during the three-dimensional survey, and is expressed as: in, Indicates the first The three-dimensional survey waypoint to the first Flight distance of each three-dimensional waypoint; This indicates the penalty weight corresponding to the energy consumption of a flight segment; both flight distance and segment energy consumption are included in the heuristic factor. The shorter the distance and the lower the energy consumption, the larger the heuristic factor, and the higher the probability that the corresponding flight segment will be selected in the ant colony search. The intensity of the energy consumption penalty can be adjusted according to the post-disaster wind field intensity, battery remaining capacity, or mission urgency, so that route planning neither detours excessively nor ignores the additional power consumption risk caused by strong winds after the typhoon.

[0068] Under the constraint of UAV operation, the access order of 3D patrol waypoints for each UAV is iteratively optimized to generate UAV patrol tasks. Each UAV patrol task includes the UAV number, the target to be patrolled, the access order of 3D patrol waypoints, and the corresponding photography task. For multiple adjacent high-risk patrol nodes within the same area, they can be merged into a single continuous patrol task by the same UAV; for high-risk patrol nodes with large spatial distances or significant differences in wind field conditions, they can be assigned to different UAVs to reduce single-aircraft range pressure and task waiting time.

[0069] Based on the UAV patrol mission, the takeoff point, 3D patrol waypoints, hovering shooting points, and return point of each UAV are sequentially connected to generate a UAV patrol route. The UAV patrol route is a sequence of 3D flight waypoints including latitude and longitude, altitude, visit order, shooting attitude, and the executing UAV number. In this way, the route results not only guide UAV flight but also clearly identify the shooting object and shooting action corresponding to each waypoint, facilitating subsequent image recognition results and backtracking matching with key patrol targets.

[0070] Unmanned aerial vehicle (UAV) operation constraints include battery power constraints, flight distance constraints, flight time constraints, weather safety constraints, airspace safety constraints, flight altitude constraints, multi-UAV collaboration constraints, return-to-home constraints, and mission coverage constraints.

[0071] Among them, the power constraint is used to limit the sum of the energy consumption of each UAV when performing the corresponding UAV patrol mission, the energy consumption of the flight segment, the energy consumption of hovering, and the energy consumption of returning to home, to not exceed the available power of the UAV. The power constraint is expressed as: in, Indicates the first The drone patrol route corresponding to each drone; This refers to a segment formed by two adjacent three-dimensional waypoints in the drone's patrol route; Indicates the first The hovering energy consumption required for a drone to perform hovering photography mission; Indicates the first The energy consumption required for a drone to return to its landing point after completing an inspection; Indicates the first The available battery power of the drone; This represents the power safety margin coefficient; Flight distance and flight time constraints are used to limit the total flight distance and total flight time of each UAV's patrol route to not exceeding the corresponding UAV's maximum allowable flight distance and maximum allowable flight time. This constraint can prevent a single UAV from undertaking excessively long routes and reduce mission interruptions caused by insufficient endurance.

[0072] Meteorological safety constraints are used to limit drones from performing patrol missions within meteorological areas that meet safe flight conditions. These constraints include wind speed not exceeding the maximum permissible wind speed, precipitation not exceeding the maximum permissible precipitation, and visibility not falling below the minimum permissible visibility. These constraints prevent drones from entering areas with excessively strong winds or excessively low visibility, improving flight safety and image acquisition quality.

[0073] Airspace safety constraints are used to limit drone patrol routes to avoid no-fly zones, temporarily controlled areas, and areas with obstacles, and to maintain a preset safe distance between drones and poles, power lines, buildings, and other drones. These constraints reduce the risk of drone collisions and prevent patrol missions from being unable to be carried out due to airspace conflicts.

[0074] Flight altitude constraints are used to limit the flight altitude of each 3D inspection waypoint to within the allowable range based on the equipment installation height, the type of equipment to be inspected, and the imaging requirements in the equipment log data. For example, for tower-type targets, the drone needs to be within a height range that allows it to photograph crossarms, insulators, and conductor connections; for line corridor-type targets, both the conductor corridor and surrounding tree obstructions need to be considered.

[0075] Multi-UAV collaborative constraints limit the assignment of a single high-risk patrol node to at most one UAV in the same patrol cycle, and restrict different UAVs from entering the same spatial conflict zone within the same time period. This constraint reduces redundant patrols and aerial conflicts, thereby improving the efficiency of multi-UAV collaborative patrols.

[0076] Return-to-home constraints ensure that each drone, after completing its assigned patrol mission, can return from the last 3D waypoint to the pre-defined landing point or emergency landing point. In areas with unstable communication or rapidly changing wind fields after a typhoon, return-to-home constraints guarantee the safe recovery of drones after their missions are completed.

[0077] The task coverage constraint prioritizes drone coverage of high-risk inspection nodes where the drone inspection urgency index reaches a preset urgency threshold. When the number of drones or their battery power is insufficient, drone inspection tasks are determined according to the drone inspection urgency index from highest to lowest. This constraint ensures that inspection resources are prioritized for critical targets most likely to affect power restoration, preventing limited drone resources from being prematurely occupied by low-risk areas.

[0078] S5. Control the drone to collect target images according to the drone's patrol route, and perform image enhancement and image recognition on the target images to obtain equipment damage identification results; In one specific implementation, the equipment damage identification results include the location of the damaged equipment, the damaged part, and the damage type. After the UAV arrives at the 3D inspection waypoint corresponding to the key inspection target according to the UAV inspection route, it acquires images of the equipment or section to be inspected based on the hovering shooting position and shooting attitude configured in the 3D inspection waypoint, thus obtaining the target image. For different objects such as towers, insulators, conductors, and hardware, the shooting angle can be adjusted according to the equipment installation height and suspected fault type in the equipment ledger data to ensure that the target image covers vulnerable parts as much as possible.

[0079] After data acquisition, the target image is correlated with the acquisition time, drone location, shooting posture, key inspection targets, and equipment ledger data to obtain the target image to be identified. Through this correlation, the subsequently identified damaged parts can be traced back to specific equipment identification, line section, and geographical location, avoiding the situation where only image-level recognition results are obtained and cannot be directly used for emergency repair dispatch.

[0080] Image degradation status is assessed before identifying the target image. Image degradation status includes at least one of the following: low illumination, heavy fog, heavy rainfall, waterlogging reflection, motion blur, and occlusion interference. Post-typhoon images are often affected by rain, fog, water reflection, fallen trees, and drone shaking. If directly input into the target detection model, minor defects such as tower cracks, broken conductor strands, and damaged insulator skirts may be missed. Therefore, image degradation status is assessed before identification, and enhancement processing can be selected based on image quality.

[0081] When the target image to be identified exhibits image degradation, anti-degradation image enhancement processing is performed on the target image at the UAV edge computing node or emergency repair command platform to obtain an enhanced target image. Anti-degradation image enhancement processing includes global restoration processing based on generative image enhancement and local contrast enhancement processing based on adaptive histograms. When the target image to be identified does not exhibit image degradation, it can be directly used as the enhanced target image to avoid unnecessary enhancement that could cause image detail distortion.

[0082] Degradation-resistant image enhancement processing includes conditional generative adversarial network (GAN) enhancement and contrast-limited adaptive histogram equalization (HXEM). The GAN consists of a generator network and a discriminator network. The generator network maps the degraded target image to a sharp image, while the discriminator network distinguishes between the real sharp image and the enhanced image generated by the generator network. The adversarial loss function of the GAN can be expressed as: in, Represents a generator network; This represents the discriminator network; This represents the adversarial loss function between the generator network and the discriminator network. This indicates the target image to be identified that has undergone image degradation. This represents the conditional or noise variables input to the generator network; This indicates that the generator network identifies the target image based on... and condition variables or noise variables The generated enhanced image; Represents a true and clear image; This indicates that the discriminator network will display a true and clear image. The probability of classifying it as a real image; This indicates that the discriminator network will enhance the image. The probability of classifying it as a real image; This represents the expectation operation on the distribution of a true, sharp image; This represents the expected operation of the joint distribution of the target image to be identified and the condition variable or noise variable.

[0083] Through adversarial training between the generator network and the discriminator network, the generator network gradually learns the mapping relationship from rainy, foggy, low-light, or reflective images to clear images.

[0084] After the generator network outputs the enhanced image, it undergoes contrast-constrained adaptive histogram equalization. This process divides the image into multiple local regions, restricts and redistributes the grayscale distribution of each region, enhances local textures such as insulator skirts, hardware bolts, tower connections, and conductor edges, while suppressing noise amplification issues that may arise from ordinary histogram equalization. The enhanced target image is obtained after this processing.

[0085] The enhanced target image is input into the damaged area recognition model to detect power grid equipment, equipment components, and damaged areas within the enhanced target image, yielding equipment detection boxes, damaged area detection boxes, damage type, and recognition confidence scores. The damaged area recognition model can employ a single-stage target detection model incorporating a collaborative attention mechanism. This single-stage model includes a backbone feature extraction network, a collaborative attention module, and a detection head. The backbone feature extraction network extracts deep convolutional feature maps from the enhanced target image, the collaborative attention module enhances feature responses related to power grid equipment and damaged areas, and the detection head outputs the target location, category, and confidence score.

[0086] The collaborative attention module includes a channel attention module and a spatial attention module. The channel attention module generates channel attention weights based on the global average pooling and global max pooling results of the deep convolutional feature maps. These channel attention weights are expressed as: in, Representing deep convolutional feature maps The corresponding channel attention weights; This represents the Sigmoid activation function; Indicates a shared multilayer perceptron; Represents the feature map of deep convolution. Channel description features obtained by performing global average pooling; Represents the feature map of deep convolution. The channel description features are obtained by performing global max pooling. Global average pooling is used to characterize the overall response of each channel, global max pooling is used to preserve strong response features, and a shared multilayer perceptron is used to learn the importance of each channel. With this setup, the model can enhance effective feature channels such as insulator contours, metal component textures, and conductor edges, while suppressing background interference channels such as water accumulation, mud, and tree branches.

[0087] The deep convolutional feature map is reconstructed using the channel attention weights to obtain the reconstructed feature map. The spatial attention module then generates spatial attention weights based on the average pooling and max pooling results of the reconstructed feature map. The spatial attention weights are expressed as follows: in, Representing deep convolutional feature maps The corresponding spatial attention weights; Indicates the kernel size as Convolution operation; This indicates that the average pooling and max pooling results will be concatenated along the channel dimension. This setting allows the model to focus more intently on locally damaged areas such as tower cracks, broken conductor strands, vibration damper displacement, and insulator breakage, while reducing interference from tree obstructions and water reflections on the recognition results.

[0088] Spatial reconstruction of the channel reconstruction feature map is performed based on spatial attention weights to obtain an attention-enhanced feature map. This attention-enhanced feature map is then input into the detection head to identify damaged regions in the enhanced target image, and outputs the device detection box, damaged area detection box, damage type, and recognition confidence score. Damage types include at least one of the following: insulator breakage, insulator string fracture, tower deformation, tower cracks, conductor strand breakage, missing hardware, and vibration damper displacement.

[0089] Based on the equipment detection frame, damaged area detection frame, recognition confidence level, drone location, shooting posture, and equipment ledger data, the location, damaged area, and damage type of the damaged equipment are determined, generating equipment damage identification results. By matching the image recognition results with the drone's positioning information, shooting posture, and equipment ledger data, the damaged area in the image can be converted into structured fault information on specific equipment or line sections, facilitating subsequent fault verification, repair dispatch, and resource allocation by the emergency repair command platform.

[0090] S6. Generate post-disaster inspection decision results based on key inspection targets, UAV inspection routes and equipment damage identification results, and transmit the post-disaster inspection decision results back to the emergency repair command platform to guide the allocation of post-disaster emergency repair resources. In one specific implementation, post-disaster inspection decision results are generated based on key inspection targets, UAV inspection routes, and equipment damage identification results. These results include damaged equipment identification, location, damaged part, damage type, identification confidence level, inspection priority, repair priority, and repair resource requirements. Damaged equipment identification and location can be jointly determined based on equipment ledger data, UAV positioning information, and image recognition results; damaged part and damage type can be provided by the equipment damage identification results; and repair priority can be determined by combining the UAV inspection high-risk urgency index, damage type severity, equipment importance, and identification confidence level.

[0091] For example, targets identified as "critical line conductor strand breakage" with a high confidence level can be assigned a higher repair priority; targets identified as "minor missing ordinary fittings" with low importance and no impact on power supply can be assigned a lower repair priority. In this way, image recognition results are no longer just annotations in inspection images, but are transformed into structured decision-making information that can be directly accessed by the emergency repair command platform.

[0092] After the post-disaster inspection decisions are transmitted back to the emergency repair command platform, the platform allocates repair personnel, vehicles, emergency supplies, power generation equipment, and drone verification tasks based on repair priorities and resource needs. Resource needs can be matched according to the type of damage; for example, broken conductor strands correspond to conductor repair materials and personnel for working at heights, deformed towers correspond to structural verification personnel and reinforcement materials, and damaged insulators correspond to spare insulator parts and arrangements for live-line or power-off work. This process reduces the time spent on secondary manual assessment and task assignment, improving the efficiency of post-disaster emergency repair command.

[0093] After the emergency repair task is completed, decision-making units are divided according to the repair responsibility area or repair grid area. Each decision-making unit can correspond to a power supply grid, repair team responsibility area, substation power supply area, or affected line section. For each decision-making unit, emergency repair input indicators and emergency repair output indicators are collected. Emergency repair input indicators include at least one of the following: manpower input time, number of repair vehicles, number of emergency lighting equipment, number of power generation equipment, amount of windproof reinforcement materials, and amount of spare parts. Emergency repair output indicators include at least one of the following: restored power load, number of key transmission sections repaired, number of equipment repaired, power restoration response speed, and number of faults eliminated. Among them, the power restoration response speed can be expressed as the reciprocal of the power restoration response time, so that areas with shorter response times have higher output values.

[0094] Based on the emergency repair input and output indicators, an input indicator matrix and an output indicator matrix are constructed. The Balanced Cost-Cut Model (BCC) from Data Envelopment Analysis (DEA) is then used to calculate the overall post-disaster recovery efficiency value for each decision-making unit. The BCC model is applicable when there are variable returns to scale between emergency repair input and recovery output. For example, different regions may have varying disaster-affected areas, equipment density, transportation conditions, and difficulties in material delivery. Even with the same resource input, different recovery effects may occur. Therefore, using the BCC model can more reasonably evaluate the relative recovery efficiency of each region.

[0095] For the The comprehensive efficiency value of post-disaster recovery for each decision-making unit is calculated using the following formula: in, Indicates the first The overall efficiency value of post-disaster recovery for each decision-making unit; Represents a non-Archimedean infinitesimal; This indicates the quantity of emergency repair inputs; This indicates the quantity of emergency repair output indicators; Indicates the number of decision-making units participating in the evaluation; Indicates the first The decision-making unit corresponds to the first... The indicator value of emergency repair investment; Indicates the first The decision-making unit corresponds to the first... The indicator value of emergency repair investment; Indicates the first The decision-making unit corresponds to the first... The indicator value of each emergency repair output indicator; Indicates the first The decision-making unit corresponds to the first... The indicator value of each emergency repair output indicator; Indicates the first The intensity coefficient corresponding to each decision-making unit when constructing the evaluation benchmark; Indicates the first Redundant slack variables corresponding to each emergency repair input indicator; Indicates the first The output insufficiency slack variable corresponding to each emergency repair output indicator; By introducing input redundancy slack variables and output insufficiency slack variables, we can not only obtain the overall efficiency value, but also further pinpoint the causes of inefficiency. For example, if a region has a large number of emergency repair vehicles but a low power restoration load, it may indicate input redundancy; if a region has a reasonable input level but a slow power restoration response speed, it may indicate insufficient output or a delayed response.

[0096] Based on the overall post-disaster recovery efficiency value, input redundancy slack variables, and output deficiency slack variables, the post-disaster recovery effectiveness evaluation results for each decision-making unit are determined. The post-disaster recovery effectiveness evaluation results include the effective recovery area, the input redundancy area, the output deficiency area, and the response lag area. When the value is close to or equal to 1, and both redundant input slack variables and insufficient output slack variables are small, the decision-making unit can be considered to have high recovery efficiency; when When the output is low, or there is obvious redundancy in input and insufficient output, it can be determined that there is room for improvement in resource allocation or emergency repair organization of the decision-making unit.

[0097] Based on the post-disaster recovery effectiveness evaluation results, disaster prevention and mitigation feedback information is generated. This feedback information includes adjustments to windproofing and reinforcement needs, adjustments to emergency repair resource allocation, adjustments to spare parts and components reserves, and optimization of material storage warehouse locations. For example, for areas experiencing repeated response delays, the pre-positioning level of emergency repair forces can be increased; for areas with redundant material input but low recovery output, the types and quantities of materials allocated can be reassessed; and for areas where critical equipment is repeatedly damaged, the level of windproofing and reinforcement can be increased or the management plan for power line corridors can be optimized.

[0098] The feedback information on disaster prevention and mitigation is fed back into pre-disaster prevention and control planning and subsequent typhoon disaster inspection decisions. This is used to update prior disaster information, drone operation constraints, emergency repair resource requirements, and rules for determining key inspection targets. Through this closed-loop process, the results of post-disaster inspections and repairs can be used to refine subsequent typhoon prevention and control experience. This allows the system to not only complete a single post-disaster inspection and repair scheduling but also continuously improve target selection, flight path planning, and resource allocation strategies for the next typhoon disaster.

[0099] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0100] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent UAV patrol decision-making after typhoon disaster based on image recognition and semantic analysis, characterized in that, Includes the following steps: S1. Acquire multi-source post-typhoon data and preprocess the multi-source post-typhoon data to obtain multi-modal post-typhoon basic data; wherein, the multi-source post-typhoon data includes meteorological forecast and measured data, equipment operation data, equipment ledger data, disaster reporting and emergency repair text data, and post-typhoon image data collected by UAVs. S2. Perform semantic analysis on the disaster reporting and emergency repair text data, extract typhoon characteristic information, affected equipment information and prone fault information related to typhoon disaster, and generate disaster prior information based on the typhoon characteristic information, the affected equipment information and the prone fault information; S3. The disaster prior information is integrated and analyzed with the equipment operation data to determine the suspected damage status of each piece of equipment or section to be inspected, and the key inspection targets are determined based on the suspected damage status and the equipment ledger data. S4. Based on the key inspection targets, the equipment ledger data, the meteorological forecast and measured data, and the drone operation constraints, generate drone inspection tasks and corresponding drone inspection routes. S5. Control the UAV to collect target images according to the UAV patrol route, and perform image enhancement and image recognition on the target images to obtain equipment damage identification results; wherein, the equipment damage identification results include the location of the damaged equipment, the damaged part and the damage type; S6. Generate post-disaster inspection decision results based on the key inspection targets, the UAV inspection route, and the equipment damage identification results, and send the post-disaster inspection decision results back to the emergency repair command platform to guide the allocation of post-disaster emergency repair resources.

2. The intelligent decision-making method for typhoon disaster post-disaster UAV patrol based on image recognition and semantic analysis as described in claim 1, characterized in that, The preprocessing of the post-disaster multi-source data to obtain post-disaster multimodal basic data specifically includes: The meteorological forecast and measured data are processed into grids and time-corrected to extract meteorological features corresponding to each piece of equipment or section to be inspected; wherein, the meteorological features include precipitation, temperature, humidity, typhoon intensity, typhoon path, wind speed and wind direction; Physical boundary screening is performed on the equipment operation data to remove abnormal data that exceeds the physical range of equipment operation, and spatiotemporal completion is performed on missing data caused by communication interruption or collection anomaly to obtain preprocessed equipment operation data. For short-term missing data in the time dimension, interpolation is performed using equipment operation data from adjacent time points to complete the missing data. Isolated anomaly detection is performed on the equipment operation data after spatiotemporal completion to remove isolated noise data generated by false alarms from post-disaster sensors, resulting in denoised equipment operation data; The equipment ledger data is subjected to unified equipment identification, standardized spatial location, and power grid topology verification to obtain preprocessed equipment ledger data; The disaster reporting and emergency repair text data are cleaned, the equipment name is standardized, the location name is standardized, and the timestamp is corrected. The standardized equipment name and location name are then matched with the preprocessed equipment ledger data for fuzzy matching and key-value mapping to obtain text intelligence data corresponding to the ledger equipment. Video frame extraction is performed on the post-disaster image data collected by the UAV to obtain candidate image frames, and the sharpness of the candidate image frames is evaluated based on the Laplace variance. Candidate image frames with a sharpness evaluation value lower than a preset sharpness threshold are identified as invalid image frames and removed, while valid post-disaster image data are retained; Using the preprocessed equipment ledger data as the correlation benchmark, the meteorological features, the denoised equipment operation data, the text intelligence data, and the effective post-disaster image data are correlated according to equipment identification, spatial location, and minute-level time information to obtain the post-disaster multimodal basic data.

3. The intelligent decision-making method for typhoon disaster relief using unmanned aerial vehicles (UAVs) based on image recognition and semantic analysis according to claim 1, characterized in that, The semantic analysis of the disaster reporting and emergency repair text data, extracting typhoon characteristic information, affected equipment information, and prone-to-fault information related to the typhoon disaster, specifically includes: The disaster reporting and emergency repair text data are segmented into sentences, invalid characters are removed, equipment names are standardized, fault description words are standardized, and place names are standardized to obtain a standardized text sequence. The normalized text sequence is input into a pre-trained language representation model, and dynamic word vector representations incorporating contextual semantics are obtained through self-attention computation. The language representation model is a BERT model built on a Transformer encoder. The BERT model includes a word embedding layer, a position embedding layer, a fragment embedding layer, and a multi-layer Transformer encoding layer, which are used to generate dynamic word vector representations based on the contextual relationships of each word in the normalized text sequence. The dynamic word vector representation is input into the bidirectional sequence feature extraction model to extract the forward context features and backward context features of the normalized text sequence, respectively. The forward context features and backward context features are then concatenated to obtain a comprehensive semantic feature vector. The bidirectional sequence feature extraction model is a BiLSTM model built on a bidirectional long short-term memory network. The BiLSTM model includes a forward long short-term memory network and a backward long short-term memory network. The forward long short-term memory network is used to extract the semantic dependencies from front to back in the normalized text sequence, and the backward long short-term memory network is used to extract the semantic dependencies from back to front in the normalized text sequence. The comprehensive semantic feature vector is input into the conditional random field decoding layer, and the normalized text sequence is globally labeled according to the label transition relationship to obtain a text label sequence; wherein, the text label sequence includes typhoon feature labels, equipment labels, component labels, fault labels, handling labels, and location labels; The conditional random field decoding layer aims to maximize the conditional probability of the text label sequence, and the conditional probability is given by the formula: in, Indicates the input text sequence Output text label sequence under certain conditions The conditional probability; This represents the normalized text sequence; This represents the sequence of text tags to be decoded; Represents any sequence of candidate text labels; Indicates the first The comprehensive semantic feature vector corresponding to each word element; Represents the text label sequence The Middle The tags corresponding to each word element; Represents the candidate text label sequence The Middle The tags corresponding to each word element; Representation and Label The corresponding feature weight parameters; Representation and Label The corresponding feature weight parameters; Indicates by label Transfer to label Tag transfer parameters; Indicates by label Transfer to label Tag transfer parameters; Based on the text tag sequence, extract typhoon characteristic entities, equipment entities, component entities, and fault entities from the normalized text sequence; Based on the semantic relationships between the typhoon characteristic entity, the equipment entity, the component entity, and the fault entity in the same disaster record or emergency repair record, a structured triple is generated; wherein, the structured triple includes typhoon characteristics, affected equipment, and prone faults; Based on the structured triples, the typhoon characteristic information, the affected equipment information, and the prone fault information are obtained.

4. The intelligent typhoon disaster relief and decision-making method for UAV patrols based on image recognition and semantic analysis according to claim 3, characterized in that, Disaster prior information is generated based on the typhoon characteristic information, the affected equipment information, and the prone-to-fault information, specifically including: Based on the typhoon characteristic information, the affected equipment information, and the prone fault information, a structured triplet set consisting of typhoon characteristics, affected equipment, and prone faults is constructed. The affected equipment in the structured triplet set is matched with the equipment ledger data by equipment identifier and spatial location to determine the equipment or section to be inspected corresponding to each affected equipment. Based on the correspondence between typhoon characteristics, equipment or sections to be inspected, and common faults, the conditional co-occurrence frequency between different entities is statistically analyzed, and a prior knowledge matrix is ​​generated. The elements in the prior knowledge matrix are represented as follows: in, Represents the first in the prior knowledge matrix The fault type corresponding to the equipment or section to be inspected. The prior probability of historical vulnerability; Represents a set of fault types; In the set of structured triples, the first... The equipment or section to be inspected and the type of fault. The number of times conditional co-occurrence occurs; This represents the Laplace smoothing coefficient, and ; Represents the set of fault types The number of fault types; Based on the prior knowledge matrix, determine the historical vulnerability prior probability and the corresponding common fault type of each equipment or section to be inspected; The prior probability of historical vulnerability, the type of common failure, and the typhoon characteristics corresponding to the type of common failure are combined to generate the prior disaster information; wherein, the prior disaster information includes the equipment or section to be inspected, the prior probability of historical vulnerability, the typhoon characteristics, and the type of common failure.

5. The intelligent decision-making method for typhoon disaster post-disaster drone patrol based on image recognition and semantic analysis as described in claim 1, characterized in that, S3 specifically includes: The historical vulnerability prior probability, typhoon characteristics, and prone failure types corresponding to each equipment or section to be inspected are obtained from the disaster prior information. The electrical abnormality signals corresponding to each piece of equipment or section to be inspected are extracted from the equipment operation data to form an electrical abnormality signal set; wherein, the electrical abnormality signals include line impedance abnormality, protection action signal, power outage alarm signal, telemetry over-limit signal, current abnormality signal, voltage abnormality signal and communication interruption signal; Based on the power grid topology connection relationship between the equipment or sections to be inspected, a Bayesian network is constructed to represent the correlation between disaster prior information, electrical anomaly signals and equipment physical damage events; Based on the Bayesian network, the prior probability of historical vulnerability is jointly inferred with the set of electrical anomaly signals to determine the posterior probability of damage for each piece of equipment or section to be inspected. The posterior probability of damage is given by the following formula: in, Indicates the electrical abnormality signal set Conditions, No. The posterior probability of physical damage to a piece of equipment or section to be inspected; Indicates the first An event in which physical damage occurs to the equipment or section to be inspected; This represents the set of electrical anomaly signals extracted from the operating data of the equipment; Indicates the first The conditional probability of the electrical abnormality signal set occurring when a piece of equipment or section to be inspected suffers physical damage; This refers to the set of equipment or sections to be inspected that are participating in the joint inference; Indicates the first The prior historical vulnerability probability corresponding to each piece of equipment or section to be inspected is expressed as: in, Represents the first in the prior knowledge matrix The fault type corresponding to the equipment or section to be inspected. The prior probability of historical vulnerability; Indicates the first A set of common fault types corresponding to each piece of equipment or section to be inspected; Indicates the first The fault type corresponding to the equipment or section to be inspected. Fault association weights; Based on the posterior damage probability and in conjunction with the common fault types in the prior disaster information, the suspected damage status of each piece of equipment or section to be inspected is determined; wherein, the suspected damage status includes the suspected damage probability, the suspected damage level, and the suspected fault type; Based on the equipment ledger data, the importance of each piece of equipment or section to be inspected is determined. Then, based on the posterior probability of damage and the equipment importance, a high-risk urgency index for drone inspection of each piece of equipment or section to be inspected is determined. The formula for the high-risk urgency index for drone inspection is: in, Indicates the first The urgency index of high-risk drone inspection of individual equipment or sections to be inspected; Indicates the first The importance of each piece of equipment or section to be inspected; This represents the weighting coefficient corresponding to the posterior probability of damage; This represents the weighting coefficient corresponding to the importance of the device; Equipment or sections to be inspected that reach a preset urgency threshold for UAV inspection high-risk urgency index are identified as candidate high-risk targets. These candidate high-risk targets are then merged and deduplicated based on their spatial location, power grid topology connection relationship, and suspected fault type to obtain key inspection targets. The key inspection targets include target equipment identification, target spatial location, suspected damage level, suspected fault type, and UAV inspection high-risk urgency index.

6. The intelligent decision-making method for typhoon disaster relief using unmanned aerial vehicles (UAVs) based on image recognition and semantic analysis according to claim 1, characterized in that, The process of generating drone inspection tasks and corresponding drone inspection routes based on the key inspection targets, the equipment ledger data, the meteorological forecast and measured data, and drone operation constraints specifically includes: Based on the target equipment identification and target spatial location in the key inspection targets, the corresponding equipment type, line, section, power grid topology connection relationship and equipment installation height are matched in the equipment ledger data to generate a set of high-risk inspection nodes; Based on the target spatial location, equipment installation height, suspected fault type, and UAV patrol urgency index of each high-risk patrol node in the set of high-risk patrol nodes, the corresponding three-dimensional patrol waypoints are determined; wherein, the three-dimensional patrol waypoints include waypoint latitude and longitude, flight altitude, hovering shooting position, and shooting attitude; Based on the meteorological forecasts and measured data, obtain the wind speed, wind direction, precipitation and visibility of the flight segment between any two three-dimensional patrol points, and determine the wind resistance impact of the corresponding flight segment in combination with the UAV flight direction; Based on the wind resistance effect, flight distance, and flight time, the segment energy consumption between any two three-dimensional survey waypoints is determined. The segment energy consumption is calculated using the following formula: in, Indicates that the drone is from the first The three-dimensional survey waypoint flew to the first Energy consumption of each three-dimensional waypoint patrol route; This indicates the airspeed energy consumption coefficient of the drone; Indicates the drag penalty coefficient; This indicates the speed of the drone relative to the air. This indicates the wind speed corresponding to the stated flight segment; Indicates the angle between the drone's flight direction and the wind direction; Based on the set of high-risk inspection nodes, the energy consumption of the flight segment, and the constraints of UAV operations, a multi-UAV collaborative path optimization model is constructed; wherein, the multi-UAV collaborative path optimization model aims to prioritize covering high-risk inspection nodes with a high urgency index for UAV inspections and reduce the energy consumption of the flight segment. The multi-objective ant colony algorithm is used to solve the multi-UAV cooperative path optimization model. In the nth iteration, the 1st Only one ant from the first The 3D survey waypoint was transferred to the first... The state transition probability of a 3D sighting waypoint is given by the formula: in, Indicates the first In the nth iteration Only one ant from the first The 3D survey waypoint was transferred to the first... The state transition probability of each three-dimensional waypoint; Indicates the first In the nth iteration The three-dimensional survey waypoint and the first Pheromon concentration between three-dimensional waypoints; Indicates the first The three-dimensional survey waypoint to the first Heuristic factors for three-dimensional waypoints; Indicates the first The urgency index of high-risk drone patrols corresponding to each three-dimensional patrol waypoint; Indicates the first The set of three-dimensional waypoints that an ant is allowed to access in its current state; This indicates the influencing factors corresponding to pheromone concentration; This represents the impact factor corresponding to the heuristic factor; This indicates the influencing factors corresponding to the high-risk urgency index of drone patrols; The heuristic factor is determined based on the flight distance and segment energy consumption between three-dimensional waypoints, and is expressed as: in, Indicates the first The three-dimensional survey waypoint to the first Flight distance of each three-dimensional waypoint; This indicates the penalty weight corresponding to the energy consumption of a flight segment; Under the condition of satisfying the UAV operation constraints, the access order of the three-dimensional inspection waypoints corresponding to each UAV is iteratively optimized to generate UAV inspection tasks; wherein, the UAV inspection task includes the UAV number, the target to be inspected, the access order of the three-dimensional inspection waypoints and the corresponding shooting task. According to the UAV patrol mission, the take-off point, three-dimensional patrol waypoint, hovering shooting point and return point of each UAV are connected in sequence to generate the UAV patrol route; wherein, the UAV patrol route is a three-dimensional flight waypoint sequence including latitude and longitude, altitude, access order, shooting attitude and execution UAV number.

7. The intelligent decision-making method for typhoon disaster post-disaster drone patrol based on image recognition and semantic analysis as described in claim 6, characterized in that, The constraints on drone operations include: The power constraint limits the sum of the energy consumption of each UAV during its patrol mission (segment energy consumption, hovering energy consumption, and return energy consumption) to no more than the available power of that UAV. The power constraint is expressed as follows: in, Indicates the first The drone patrol route corresponding to each drone; This refers to a segment formed by two adjacent three-dimensional waypoints in the UAV's patrol route; Indicates the first The hovering energy consumption required for a drone to perform hovering photography mission; Indicates the first The energy consumption required for a drone to return to its landing point after completing an inspection; Indicates the first The available battery power of the drone; This represents the power safety margin coefficient; Flight distance constraints and flight time constraints are used to limit the total flight distance and total flight time of each UAV corresponding to the UAV patrol route to not exceed the maximum allowable flight distance and maximum allowable flight time of the corresponding UAV. Meteorological safety constraints are used to limit the UAV from performing patrol missions within meteorological areas that meet safe flight conditions; wherein, the meteorological safety constraints include wind speed not exceeding the maximum permissible wind speed, precipitation not exceeding the maximum permissible precipitation, and visibility not being lower than the minimum permissible visibility; Airspace safety constraints are used to restrict the UAV's patrol route from avoiding no-fly zones, temporary control zones, and obstacle zones, and to maintain a preset safe distance between the UAV and poles, power lines, buildings, and other UAVs; Flight altitude constraints are used to limit the flight altitude of each three-dimensional inspection waypoint to within the allowable altitude range based on the equipment installation height, the type of equipment to be inspected, and the shooting requirements in the equipment ledger data. Multi-drone collaborative constraints are used to limit the same high-risk inspection node to a maximum of one drone in the same inspection round, and to restrict different drones from entering the same spatial conflict area within the same time period; Return-to-home constraints are used to ensure that each UAV can return to the preset landing point or emergency landing point from the last three-dimensional waypoint after completing its corresponding UAV inspection mission. The task coverage constraint is used to prioritize drone coverage of high-risk inspection nodes where the drone inspection urgency index reaches a preset urgency threshold, and to determine drone inspection tasks to be executed in descending order of the drone inspection urgency index when the number of drones or battery power is insufficient.

8. The intelligent patrol and decision-making method for typhoon disaster relief using unmanned aerial vehicles based on image recognition and semantic analysis according to claim 1, characterized in that, S5 specifically includes: The drone is controlled to reach the three-dimensional inspection waypoint corresponding to the key inspection target according to the drone inspection route, and the image is acquired based on the hovering shooting position and shooting posture corresponding to the three-dimensional inspection waypoint to obtain the target image; The target image is associated with the corresponding acquisition time, drone location, shooting posture, key patrol targets, and equipment ledger data to obtain the target image to be identified; The image to be identified is subjected to an image degradation state judgment; wherein, the image degradation state includes low light, heavy fog, heavy rain, water flooding reflection, motion blur and occlusion interference; When the target image to be identified is in a state of image degradation, anti-degradation image enhancement processing is performed on the target image to be identified at the edge computing node of the UAV or the emergency repair command platform to obtain an enhanced target image; wherein, the anti-degradation image enhancement processing includes global restoration processing based on generative image enhancement and local contrast enhancement processing based on adaptive histogram. The enhanced target image is input into the damaged part recognition model to detect the power grid equipment, equipment components and damaged areas in the enhanced target image, and to obtain the equipment detection box, the damaged part detection box, the damage type and the recognition confidence. Based on the device detection frame, the damaged part detection frame, the recognition confidence level, the drone position, the shooting posture, and the device ledger data, the location of the damaged device, the damaged part, and the damage type are determined, and the device damage recognition result is generated.

9. A method for intelligent UAV patrol and decision-making based on image recognition and semantic analysis after a typhoon disaster, as described in claim 8, is characterized in that... The anti-deterioration image enhancement processing and the damaged area recognition model specifically include: The anti-deterioration image enhancement processing includes conditional generative adversarial network enhancement processing and contrast-limited adaptive histogram equalization processing. The conditional generative adversarial network includes a generator network and a discriminator network. The generator network is used to map a target image to be identified that has deteriorated into a clear image. The discriminator network is used to distinguish between a real clear image and an enhanced image generated by the generator network. The enhanced image is subjected to contrast-limited adaptive histogram equalization processing. By limiting and redistributing the gray-level distribution of local image regions, the local texture of insulator skirts, hardware bolts, tower connections and conductor edges is enhanced to obtain the enhanced target image. The damaged area identification model is a single-stage target detection model that incorporates a collaborative attention mechanism. The single-stage target detection model includes a backbone feature extraction network, a collaborative attention module, and a detection head. The enhanced target image is input into the backbone feature extraction network to extract deep convolutional feature maps; The collaborative attention module includes a channel attention module and a spatial attention module; The channel attention module generates channel attention weights based on the global average pooling and global max pooling results of the deep convolutional feature maps. Based on the channel attention weights, the deep convolutional feature map is reconstructed to obtain a channel-reconstructed feature map. The spatial attention module generates spatial attention weights based on the average pooling and max pooling results of the channel reconstructed feature maps; Based on the spatial attention weights, the channel reconstruction feature map is spatially reconstructed to obtain an attention-enhanced feature map. The attention-enhanced feature map is input into the detection head to identify the damaged area in the enhanced target image, and the device detection box, the damaged part detection box, the damage type, and the recognition confidence are output; wherein, the damage type includes insulator damage, insulator string breakage, tower deformation, tower crack, conductor strand breakage, missing hardware, and vibration damper displacement.

10. A method for intelligent UAV patrol and decision-making based on image recognition and semantic analysis after a typhoon disaster, as described in claim 1, is characterized in that... S6 specifically includes: Based on the key inspection targets, the UAV inspection route, and the equipment damage identification results, a post-disaster inspection decision result is generated; wherein, the post-disaster inspection decision result includes damaged equipment identification, damaged equipment location, damaged part, damage type, identification confidence level, inspection priority, emergency repair priority, and emergency repair resource requirements; The post-disaster inspection decision results are transmitted back to the emergency repair command platform, which then allocates emergency repair personnel, vehicles, emergency supplies, power generation equipment, and drone verification tasks according to the emergency repair priority and the emergency repair resource requirements. After the emergency repair task is completed, decision-making units are divided according to the emergency repair responsibility area or emergency repair grid area, and the emergency repair input indicators and emergency repair output indicators corresponding to each decision-making unit are collected. The emergency repair input indicators include the time spent on manual labor, the number of emergency repair vehicles, the number of emergency lighting devices, the number of power generation devices, the amount of windproof and reinforcement materials invested, and the amount of spare parts invested. The emergency repair output indicators include the restoration of power load, the number of key transmission sections repaired, the number of equipment repaired, the power restoration response speed, and the number of faults eliminated. Based on the emergency repair input indicators and the emergency repair output indicators, an input indicator matrix and an output indicator matrix are constructed, and the BCC model in data envelopment analysis is used to calculate the comprehensive post-disaster recovery efficiency value of each decision-making unit. For the For each decision-making unit, the comprehensive efficiency value of post-disaster recovery is calculated according to the following formula: in, Indicates the first The overall efficiency value of post-disaster recovery for each decision-making unit; Represents a non-Archimedean infinitesimal; This indicates the quantity of emergency repair inputs; This indicates the quantity of emergency repair output indicators; Indicates the number of decision-making units participating in the evaluation; Indicates the first The decision-making unit corresponds to the first... The indicator value of emergency repair investment; Indicates the first The decision-making unit corresponds to the first... The indicator value of emergency repair investment; Indicates the first The decision-making unit corresponds to the first... The indicator value of each emergency repair output indicator; Indicates the first The decision-making unit corresponds to the first... The indicator value of each emergency repair output indicator; Indicates the first The intensity coefficient corresponding to each decision-making unit when constructing the evaluation benchmark; Indicates the first Redundant slack variables corresponding to each emergency repair input indicator; Indicates the first The output insufficiency slack variable corresponding to each emergency repair output indicator; Based on the comprehensive post-disaster recovery efficiency value, the input redundancy slack variable, and the output deficiency slack variable, the post-disaster recovery effectiveness evaluation results of each decision-making unit are determined; wherein, the post-disaster recovery effectiveness evaluation results include the effective recovery area, the input redundancy area, the output deficiency area, and the response lag area; Based on the post-disaster recovery effectiveness evaluation results, disaster prevention and mitigation feedback information is generated; wherein, the disaster prevention and mitigation feedback information includes information on adjustments to windproofing and reinforcement needs, information on adjustments to emergency repair resource allocation, information on adjustments to spare parts reserves, and information on optimization of material reserve warehouse locations; The disaster prevention and mitigation feedback information is fed back into the pre-disaster prevention and control plan and subsequent typhoon disaster inspection decision-making, and is used to update the disaster prior information, the constraints of drone operations, the demand for emergency repair resources, and the rules for determining the key inspection targets.

Images