A method for inspection optimization and damage assessment of a hybrid tower wind farm

Abstract

The application discloses a kind of mixed tower wind farm's inspection optimization and damage assessment method, belong to wind power equipment inspection technical field.The mixed tower wind farm's inspection optimization and damage assessment method includes the following steps: using unmanned aerial vehicle to screen and photograph the tower of wind farm, collect the visible light image of mixed tower, infrared thermal image and photographing pose data;Output tower suspected damage area, damage category, confidence and the corresponding machine position number of this unmanned aerial vehicle;According to tower single tower risk score model, the structural damage risk score of each machine position is calculated;Wall climbing robot carries out close review to the suspected damage area of the tower;Ground analysis platform generates structured inspection evaluation report.The mixed tower wind farm's inspection and damage assessment method of the application solves the industry pain points that existing technology unmanned aerial vehicle and wall climbing robot operation are fragmented, risk determination relies on artificial experience, damage assessment lacks structural mechanics basis, global efficiency and quantitative detection accuracy cannot be considered.

Description

Technical Field

[0001] This invention relates to the field of wind power equipment inspection technology, specifically to a method for optimizing inspection routes and assessing damage in hybrid tower wind farms based on the collaborative work of drones and wall-climbing robots. Background Technology

[0002] Wind turbine towers are constructed by combining a lower concrete tower section, an upper steel tower section, and an intermediate transition connecting section. Due to their strong material adaptability and controllable transportation and installation costs, they have become the mainstream tower structure for large-capacity, high-hub-height wind turbines. However, during service, the three types of tower sections experience differentiated structural damage. The continuous evolution of various defects significantly weakens the tower's durability and seriously threatens the long-term operational safety of the turbine. Therefore, routine and meticulous tower inspections have become a rigid core requirement in the wind power operation and maintenance field.

[0003] The tower is subjected to multiple effects over a long period of time, including wind-induced alternating fatigue loads, erosion from alternating temperature and humidity environments, and mechanical differences at the steel-concrete interface. The damage mechanisms and defect types differ among different structural sections: concrete tower sections are prone to damage such as multi-directional cracks, surface spalling, and internal water seepage; steel tower sections generally suffer from hidden dangers such as coating damage, steel corrosion, weld failure, and localized temperature anomalies; and transition connection sections, as critical load-bearing joints, are prone to structural defects such as flange misalignment, bolt loosening, and interface water seepage and corrosion.

[0004] Existing methods for inspecting wind turbine towers include drone aerial inspection and rope-assisted wall-climbing robot wall-feeding inspection. Although drones and wall-climbing robots can respectively achieve wide-area image acquisition and close-range precise measurement, they can theoretically complement each other to support the comprehensive inspection of tower defects. Long-distance drone aerial photography can qualitatively identify suspected damage, while rope-assisted wall-climbing robots can complete millimeter-level damage measurement. In actual inspection processes, the above methods are used in combination.

[0005] However, the existing inspection system has not formed an integrated operational logic that connects the entire area for initial screening, detailed review, structural safety classification, and full-cycle data traceability. The two types of intelligent inspection equipment operate independently, and it is difficult to unify the standards for identifying and judging various types of damage. It can only rely on the single dimension of appearance and size to judge the degree of damage risk, and cannot combine the actual stress state of the tower to comprehensively assess the structural risk. The task arrangement for damage review relies entirely on subjective human judgment, lacking quantitative sorting basis. Furthermore, damage data collected by different equipment and different inspection cycles cannot be automatically correlated and matched. At the same time, there is a lack of differentiated damage assessment standards for concrete, steel, and transition connection sections. Inspection information collected from multiple channels is difficult to integrate into a standardized and complete assessment conclusion that can directly guide operation and maintenance. Ultimately, this results in an imbalance in the allocation of inspection resources, high-risk towers are prone to delayed or missed inspections, and the objectivity of damage risk assessment results is insufficient. The overall inspection assessment efficiency and the completeness of safety management are both lacking. Summary of the Invention

[0006] The purpose of this invention is to overcome the problems in the prior art and provide an optimized inspection and damage assessment method for hybrid tower wind farms. This method solves the core pain points of the industry, such as the disconnect between the operation of drones and wall-climbing robots, the reliance on human experience for risk assessment, the lack of structural mechanics basis for damage assessment, and the inability to balance overall efficiency and quantitative detection accuracy.

[0007] This invention provides a method for optimizing the inspection and assessing the damage of hybrid-tower wind farms, comprising the following steps: Drones were used to screen and photograph the towers of wind farms, and to collect screening data of the towers. The screening data of the tower is input into the edge computing module of the UAV for rapid screening of end-side damage, and the suspected damage area of ​​the tower, damage category, confidence level and the corresponding UAV position number are output. Based on the suspected damage area, damage type, confidence level, and the corresponding drone position number, computer vision algorithms are used to analyze the area ratio of the suspected damage area of ​​the tower, crack width, infrared temperature difference, spatial location of the suspected damage, tower segment type, and structural sensitivity indicators based on the images taken by the drone. Based on the finite element numerical model of a single tower, the stress and dynamic response indices of each tower structure are obtained through simulation calculation. Based on the degree of structural damage simulated, the structural damage risk score of each tower is calculated, and a priority queue for review of the climbing robot is generated according to the structural damage risk score. Based on the review priority queue, combined with the current position of the wall-climbing robot, the road distance matrix of the wind farm, the endurance status of the wall-climbing robot and the single tower operation time constraint, a cross-tower review scheduling scheme for the wall-climbing robot is generated. The wall-climbing robot reaches the tower with suspected damage according to the cross-tower verification scheduling scheme, performs close-range verification of the suspected damage area of ​​the tower, and collects verification data of the suspected damage area; The ground analysis platform inputs the screening data, verification data, aircraft position information, tower segment type and UAV historical inspection records into the ground-based multimodal semantic analysis model, and generates a structured inspection assessment report by combining the structural damage grading rules.

[0008] Preferably, the rapid end-side damage screening model and the ground-based multimodal semantic analysis model are constructed as follows: Visible light images, infrared thermal images, and spatial measurement data were collected from the concrete tower section, steel tower section, and transition connection section of the tower to construct a multimodal inspection dataset containing damage category, spatial location, and damage scale annotations. A rapid end-side damage screening model and a ground-side multimodal semantic analysis model were trained based on a multimodal inspection dataset. The rapid end-side damage screening model is deployed on the edge computing module of the UAV and the rope-assisted wall-climbing robot, while the ground-based multimodal semantic analysis model is deployed on the ground analysis platform. The rapid end-side damage screening model outputs the suspected damage area, category and basic scale in real time, while the ground-based multimodal semantic analysis model is used for multi-source inspection data fusion analysis and structured damage assessment.

[0009] Preferably, the specific steps for constructing a multimodal inspection dataset are as follows: Visible light images, infrared thermographs, and spatial measurement data were collected from concrete tower sections, steel tower sections, and transition connection sections at wind farm sites or in the laboratory tower test components. Based on the collected visible light images, infrared thermal images and spatial measurement data, vertical cracks, circumferential cracks, diagonal cracks, spalling, exposed reinforcement and water seepage in the concrete tower section are marked; rust, coating damage, weld abnormalities and temperature abnormalities in the steel tower section are marked; and loose bolts, flange misalignment, water seepage marks and local rust in the transition connection section are marked. Establish a correlation record for each damage on the tower, including damage type, tower segment type, relative tower base height, circumferential angle, geometric dimensions, acquisition equipment, and acquisition time; The multimodal inspection dataset obtained after establishing the association records is organized and divided into layers according to damage type, tower segment type, and acquisition equipment.

[0010] Preferably, the rapid end-side damage screening model includes a target detection branch, a semantic segmentation branch, and an infrared anomaly discrimination branch. The target detection branch is used to output the bounding box and category of the damage. The semantic segmentation branch is used to output pixel-level masks for crack, spalling, and corrosion areas. The infrared anomaly discrimination branch is used to determine whether the temperature is abnormal based on the infrared temperature difference between the damaged area and the adjacent normal area.

[0011] Preferably, before screening and photographing the towers of the wind farm, the UAV pre-generates an inspection path. The UAV then performs screening and photographing along the inspection path. The inspection path is generated based on the spatial coordinates of each UAV position in the wind farm, the tower height, the UAV's endurance constraints, the take-off and landing point locations, and the wind farm's meteorological conditions. The screening data collected by the UAV for the towers includes visible light images, infrared thermal images, and photographing pose data of the outer surface of the tower.

[0012] Preferably, the method for collecting crack widths in the verification data is as follows: The working distance from the camera of the wall-climbing robot to the surface of the tower is determined based on laser spatial measurement data; Establish pixel scale transformation relationship by combining camera intrinsic parameters and image pixel size; The semantic segmentation mask of the crack is skeletonized, the crack pixel width is calculated along the skeleton normal, and the crack pixel width is converted into the actual crack width.

[0013] Preferably, the ground-based multimodal semantic analysis model takes damage images, segmentation masks, damage geometric measurements, infrared temperature differences, station numbers, tower segment types, relative tower base height of the damaged area, ambient wind speed, ambient temperature, and historical inspection records as inputs, and outputs an inspection assessment report in the form of structured fields. The structured fields include at least the damage name, damage area, measurement value, risk level, historical change rate, recommended re-inspection cycle, and treatment measures.

[0014] Preferably, the structural damage grading rule is classified according to different tower section types, including: For concrete tower sections, they are classified according to crack width, crack length, crack direction, spalling area, spalling depth, and water seepage status; For steel tower sections, classification is based on rust area, coating damage area, abnormal weld location, and degree of temperature abnormality; For transitional connection sections, classification is based on the amount of misalignment at the connection points, signs of bolt abnormalities, water seepage marks on the flanges, and localized corrosion.

[0015] Preferably, both the screening data from the UAV and the verification data from the wall-climbing robot contain unified spatial identification information, which includes the drone position number, tower segment type, relative tower base height, circumferential angle, and acquisition time, in order to match the records of the same damage in the UAV and the wall-climbing robot, as well as in different inspection cycles of the two.

[0016] Preferably, when generating the structured inspection and assessment report, the ground analysis platform also generates a damage distribution map and a review task record table. The damage distribution map uses the tower unfolded diagram as the base map to annotate the location, type, and risk level of each damage. The review task record table records the consistency of the UAV screening results, the wall-climbing robot review results, and the final assessment conclusion.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides an optimization method for inspection and damage assessment in hybrid tower wind farms. It integrates a heterogeneous robot hierarchical collaborative closed-loop inspection and assessment system that combines multimodal visual recognition with finite element structural mechanics simulation. First, it relies on drones to achieve rapid initial damage screening across the entire wind farm. Then, it combines apparent damage quantification indicators with tower finite element simulation stress indicators to establish a quantitative risk scoring model. Based on risk priority, it dynamically schedules wall-climbing robots to perform high-precision close-range damage verification. Finally, a standardized operation and maintenance assessment report is automatically generated by a ground-based fusion analysis model combined with differentiated damage grading rules for different tower sections. Unified spatial identification enables automatic matching and closed-loop iteration of damage data across equipment and cycles. From data source, damage visual recognition, risk scheduling, accurate measurement, to comprehensive assessment, this invention solves the core pain points of existing technologies, such as the disconnect between drone and wall-climbing robot operations, reliance on human experience for risk judgment, lack of structural mechanics basis for damage assessment, and the inability to balance overall efficiency and quantitative detection accuracy. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a flowchart illustrating the construction process of the multimodal inspection dataset for wind power concrete towers in this invention. Figure 3 This is a schematic diagram of the multi-position inspection path planning for wind farm level of the UAV according to the present invention; Figure 4 This is a schematic diagram illustrating the collaborative scheduling of the drone screening results to the wall-climbing robot verification task in this invention; Figure 5 This is a schematic diagram of the operation of the present invention; Figure 6 This is a flowchart for generating a structured inspection and evaluation report using the ground analysis platform of this invention. Detailed Implementation

[0019] The following is in conjunction with the appendix Figures 1-6 The specific embodiments of the present invention will be described in detail below, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0020] like Figures 1-6 As shown, the present invention provides an inspection optimization and damage assessment method for hybrid tower wind farms, comprising the following steps: Construct a rapid screening model for end-side injuries and a multimodal semantic analysis model for ground-based injuries; Drones were used to screen and photograph the towers of wind farms, and screening data of the towers was collected. The screening data included visible light images, infrared thermal images and shooting pose data of the outer surface of the towers. The screening data of the tower is input into the edge computing module of the UAV for rapid screening of end-side damage, and the suspected damage area of ​​the tower, damage category, confidence level and the corresponding UAV position number are output. Based on the suspected damage area, damage type, confidence level, and the corresponding drone position number, computer vision algorithms are used to analyze the area ratio of the suspected damage area of ​​the tower, crack width, infrared temperature difference, spatial location of the suspected damage, tower segment type, and structural sensitivity indicators based on the images taken by the drone. A finite element numerical model of a single tower is established, and the stress and dynamic response indices of each tower structure are obtained by simulation calculation based on the finite element numerical model of the single tower. Based on the degree of structural damage simulated, the structural damage risk score of each tower is calculated, and a priority queue for review of the climbing robot is generated according to the structural damage risk score. Based on the review priority queue, a cross-tower review scheduling scheme for the wall-climbing robot is generated by combining the current position of the wall-climbing robot, the road distance matrix of the wind farm, the endurance status of the wall-climbing robot and the single-tower operation time constraint. The wall-climbing robot reaches the tower with suspected damage according to the cross-tower verification scheduling scheme, performs close-range verification of the suspected damage area of ​​the tower, and collects verification data of the suspected damage area. The verification data includes close-range visible light images, infrared thermal images and laser spatial measurement data, crack width, crack length, peeling area, corrosion area, infrared temperature difference and misalignment of connection parts. The ground analysis platform inputs the screened data, verified data, tower location information, tower segment type, and UAV historical inspection records into the ground-based multimodal semantic analysis model. Combined with structural damage grading rules, it generates a structured inspection assessment report. The structured inspection assessment report includes the UAV location number, tower segment location, damage type, damage scale, risk level, historical trend, verified image evidence, and maintenance priority recommendations.

[0021] The working principle of the above embodiments is briefly described below: The wall-climbing robot used in this invention comprises a rope load reduction unit, an adsorption walking unit, a posture adjustment unit, a vision acquisition unit, an infrared acquisition unit, a laser ranging unit, and a wireless communication unit. The rope load reduction unit provides a safe traction force along the height of the tower. The adsorption walking unit is used for stable attachment and movement on the outer surfaces of the concrete and steel tower sections. The posture adjustment unit is used to adjust the normal angle between the vision acquisition unit and the curved surface of the tower.

[0022] Constructing a rapid end-side injury screening model and a ground-based multimodal semantic analysis model: Two types of dedicated models were trained and deployed in advance using multimodal damage samples from wind turbine concrete towers: End-side damage rapid screening model: Lightweight design, adapted to the computing power of built-in edge computing modules of drones and wall-climbing robots, can perform local real-time inference when the device acquires images, without relying on background transmission, and directly output three types of lightweight recognition results: suspected damage area, damage type, and damage basic scale, to meet the needs of rapid on-site screening. Ground-based multimodal semantic analysis model: Deployed on a ground analysis platform with sufficient computing power, it is compatible with various types of heterogeneous inspection data such as images, infrared, size, location, and historical records. It completes multi-source data fusion calculations and automatically outputs standardized structured damage assessment conclusions by matching engineering judgment criteria.

[0023] Generate an inspection path for the drone, and the drone collects tower screening data along the inspection path: Using the coordinates of all wind turbine locations, tower height, UAV endurance, environmental weather conditions, and no-fly zones as constraints, the optimal flight inspection path is solved through a multi-objective cost optimization algorithm. The UAV flies around the outside of the wind turbine concrete tower along the planned path, simultaneously collecting three types of screening data: visible light images, infrared thermal images, and shooting posture data of the tower exterior. This fully records the original appearance and temperature information of each tower, achieving wide-area full-coverage initial data collection for the entire wind farm.

[0024] The screening data is fed into the edge computing module of the drone, which outputs suspected damage information. The UAV's onboard edge computing module locally loads the collected screening data and sends it to the built-in end-side damage rapid screening model for real-time inference and calculation. The model identifies defects on the tower surface, automatically marks the outline of suspected damage areas, determines the initial damage category, and outputs the recognition confidence level. At the same time, it binds the current tower's corresponding location number and transmits the complete set of suspected damage marking information back to the ground analysis platform in real time to complete the preliminary identification of wind field-level damage.

[0025] Extracting multi-dimensional quantitative evaluation indicators for damage: After receiving suspected damage information from the UAV, the ground analysis platform performs feature quantification and extraction for each suspected damage: based on image and pose data, it calculates the area ratio of the damaged area, estimates the crack width, calculates the infrared temperature difference between the damaged area and the adjacent normal area, and simultaneously matches the drone position information to lock the spatial location of the suspected damage. It distinguishes the concrete tower segment / steel tower segment / transition connection segment to which the current damage belongs, and determines the structural sensitivity index in combination with the stress characteristics of the tower structure, forming a complete set of quantitative parameters for risk scoring.

[0026] Build a single-tower risk scoring model, calculate the risk of each tower, and generate a review priority queue: All the quantitative indicators extracted in the previous step are input into the preset single tower risk scoring model. The model calculates the overall structural damage risk score of a single tower position based on the weight of damage type, tower segment structure, and height and position. All tower positions are sorted from high to low according to the risk score value, and a priority queue for wall-climbing robot review is automatically generated to realize priority review of towers with high safety hazards.

[0027] Generate a tower-crossing review and scheduling scheme for the wall-climbing robot and distribute damage location data: Using the review priority queue as the core constraint, and superimposed with multiple constraints such as the real-time parking position of the wall-climbing robot, the road transportation distance of each turbine position in the wind farm, the remaining battery life of the equipment, and the time consumption of the review operation of a single tower, the optimal cross-tower movement sequence is solved, and a standardized cross-tower review scheduling scheme is output. The platform simultaneously sends the precise spatial coordinates of suspected damage recorded by the UAV and the visible light screening image of the tower to the control terminal of the wall-climbing robot, providing a positioning basis for the robot's precise point inspection.

[0028] The wall-climbing robot arrived on site to perform close-range verification and collect verification data: The wall-climbing robot is transported to the target position according to the scheduling plan. It moves along the outer wall of the tower by adsorption walking structure and accurately arrives at the suspected damage area according to the received damage spatial coordinates. It conducts close-range fine inspection through airborne high-definition vision, infrared thermal imaging and laser ranging equipment, and collects high-precision close-range visible light images, infrared thermal images and laser spatial measurement data. It simultaneously quantifies the measured crack length and width, peeling area, corrosion area, temperature difference and connection flange misalignment and other precise damage scales to make up for the lack of accuracy of long-distance measurement by UAV.

[0029] The ground analysis platform generates structured inspection and evaluation reports using a ground-based multimodal semantic analysis model. The ground analysis platform aggregates three types of input information: data from drone coarse screening, data from wall-climbing robots fine verification, basic information about the drone site, tower segment type, and historical inspection records. All data is uniformly sent to the ground-based multimodal semantic analysis model, which performs multimodal information fusion, comparison of new and old damage data, and calculation of damage development trends. Combining differentiated structural damage grading rules for concrete tower segments, steel tower segments, and transition connection segments, the platform classifies risk levels and ultimately outputs a structured inspection and assessment report that includes drone site number, damage area, measured damage scale, risk level, historical change trend, verification images, and maintenance priority suggestions, directly supporting operation and maintenance decisions.

[0030] in: The optimization target for the drone wind farm level inspection path is established using the following formula: In the formula, The cost of the inspection path for drones. This represents the number of machine positions to be inspected. For the first Each inspection station This refers to the flight distance between adjacent inspection stations. For the first Hovering time for each camera position The penalty items are for risks caused by wind speed, visibility, or no-fly zones. , , These are the weighting coefficients.

[0031] The single-tower risk scoring model is as follows: In the formula, For the first Structural damage risk score for each machine position For the first Number of suspected damages per camera position For the first Damage category weighting, Weights for tower segment types, Vertical position weights, This represents the ratio of the damaged area to the visible surface area of ​​the corresponding tower segment. This is a normalized value for the crack width or the misalignment at the connection point. This is the normalized value of the infrared temperature difference. This is the normalized value of the damage development rate. , , This is the dimensional compatibility factor.

[0032] Damage category weights Based on structural engineering risks, the weighting of through-cracks in concrete, misalignment in transition joints, and abnormal welds in steel tower sections is greater than that of surface spalling, localized corrosion, and coating damage; the weighting of tower section types is also considered. In the process, the weight of the transition connection section is greater than the weight of the concrete tower section and the steel tower section; the weight of the vertical position... The location is determined based on the height of the damage from the tower base and the stress-sensitive zone it occupies.

[0033] The optimization objective for the wall-climbing robot cross-tower verification scheduling scheme is established using the following formula: In the formula, To verify the cost of scheduling, This represents the number of workstations awaiting review. For the first in the verification sequence Each camera position This refers to the transport time between adjacent verification stations. For the first Estimated review time per unit For the first Structural damage risk score for each machine position As a risk priority coefficient, To avoid constants with a denominator of zero.

[0034] The structured inspection and assessment report is entered into the wind farm concrete tower damage database and used as the normalized value for calculating the damage development rate in step S4 of the next inspection cycle. The data sources enable a closed-loop update system for drone screening, robot verification, and structural damage assessment.

[0035] The present invention provides an optimized inspection and damage assessment method for hybrid-tower wind farms. This method integrates a heterogeneous robot-based hierarchical collaborative closed-loop inspection and assessment system that combines multimodal AI visual recognition with finite element structural mechanics simulation. First, it utilizes drones to achieve rapid initial damage screening across the entire wind farm. Then, it establishes a quantitative risk scoring model by combining apparent damage quantification indicators with tower finite element simulation stress indicators. Based on risk priority, it dynamically schedules wall-climbing robots to perform high-precision close-range damage verification. Finally, a standardized operation and maintenance assessment report is automatically generated by a ground-based fusion analysis model combined with differentiated damage grading rules for different tower segments. A unified spatial identifier enables automatic matching and closed-loop iteration of damage data across equipment and cycles. This method addresses the core pain points of existing technologies, such as the disconnect between drone and wall-climbing robot operations, reliance on human experience for risk assessment, lack of structural mechanics basis for damage assessment, and the inability to balance overall efficiency with quantitative detection accuracy.

[0036] As a preferred option, such as Figures 1-3 As shown, the rapid end-side injury screening model and the ground-side multimodal semantic analysis model are constructed as follows: Visible light images, infrared thermography, and spatial measurement data were collected from the concrete tower section, steel tower section, and transition connection section of the tower to construct a multimodal inspection dataset containing damage category, spatial location, and damage scale annotations. Based on the multimodal inspection dataset, an end-side damage rapid screening model and a ground-based multimodal semantic analysis model were trained. The end-side damage rapid screening model was deployed on the edge computing module of a UAV and a rope-assisted wall-climbing robot, while the ground-based multimodal semantic analysis model was deployed on a ground analysis platform. The end-side damage rapid screening model outputs the suspected damage area, category, and basic scale in real time, while the ground-based multimodal semantic analysis model is used for multi-source inspection data fusion analysis and structured damage assessment.

[0037] As a preferred option, such as Figure 1 and Figure 2 As shown, the specific steps for constructing the multimodal inspection dataset are as follows: Visible light images, infrared thermographs, and spatial measurement data were collected from concrete tower sections, steel tower sections, and transition connection sections at wind farm sites or in the laboratory tower test components. Based on the collected visible light images, infrared thermal images and spatial measurement data, vertical cracks, circumferential cracks, diagonal cracks, spalling, exposed reinforcement and water seepage in the concrete tower section are marked; rust, coating damage, weld abnormalities and temperature abnormalities in the steel tower section are marked; and loose bolts, flange misalignment, water seepage marks and local rust in the transition connection section are marked. Establish a correlation record for each damage on the tower, including damage type, tower segment type, relative tower base height, circumferential angle, geometric dimensions, acquisition equipment, and acquisition time; The multimodal inspection dataset obtained after establishing the association records is organized and divided into layers according to damage type, tower segment type, and acquisition equipment.

[0038] As a preferred option, such as Figure 1 and Figure 6 As shown, the rapid end-side damage screening model includes a target detection branch, a semantic segmentation branch, and an infrared anomaly discrimination branch. The target detection branch is used to output the bounding box and category of the damage. The semantic segmentation branch is used to output the pixel-level mask of the crack, peeling, and corrosion areas. The infrared anomaly discrimination branch is used to determine whether the temperature is abnormal based on the infrared temperature difference between the damaged area and the adjacent normal area.

[0039] As a preferred option, such as Figure 3 As shown, before screening and photographing the towers of a wind farm, the UAV pre-generates an inspection path. The UAV then performs screening and photographing along this path. The inspection path is generated based on the spatial coordinates of each UAV position within the wind farm, the tower height, the UAV's endurance constraints, the take-off and landing point locations, and the wind farm's meteorological conditions. The screening data collected by the UAV includes visible light images, infrared thermal images, and photographing pose data of the tower's outer surface. As a preferred option, such as Figure 1 As shown, the method for collecting crack widths in the verification data is as follows: The working distance from the camera of the wall-climbing robot to the surface of the tower is determined based on laser spatial measurement data; Establish pixel scale transformation relationship by combining camera intrinsic parameters and image pixel size; The semantic segmentation mask of the crack is skeletonized, the crack pixel width is calculated along the skeleton normal, and the crack pixel width is converted into the actual crack width.

[0040] As a preferred option, such as Figure 1 , Figure 2 and Figure 6As shown, the ground-based multimodal semantic analysis model takes damage images, segmentation masks, damage geometric measurements, infrared temperature differences, station numbers, tower segment types, relative height of the damaged area to the tower base, ambient wind speed, ambient temperature, and historical inspection records as inputs, and outputs an inspection assessment report in the form of structured fields. The structured fields include at least the damage name, damage area, measurement value, risk level, historical change rate, recommended re-inspection cycle, and treatment measures.

[0041] As a preferred option, such as Figure 1 and Figure 2 As shown, the structural damage grading rules are classified according to different tower section types, including: For concrete tower sections, they are classified according to crack width, crack length, crack direction, spalling area, spalling depth, and water seepage status; For steel tower sections, classification is based on rust area, coating damage area, abnormal weld location, and degree of temperature abnormality; For transitional connection sections, classification is based on the amount of misalignment at the connection points, signs of bolt abnormalities, water seepage marks on the flanges, and localized corrosion.

[0042] As a preferred option, such as Figure 1 and Figure 6 As shown, the screening data of the UAV and the verification data of the wall-climbing robot both contain unified spatial identification information. The unified spatial identification information includes the drone position number, tower segment type, relative tower base height, circumferential angle, and collection time, so as to match the records of the same damage in the UAV and the wall-climbing robot and in different inspection cycles of the two.

[0043] As a preferred option, such as Figure 1 and Figure 6 As shown, when generating a structured inspection and assessment report, the ground analysis platform also generates a damage distribution map and a review task record table. The damage distribution map uses the tower unfolded diagram as a base map to indicate the location, type, and risk level of each damage. The review task record table records the consistency of the UAV screening results, the wall-climbing robot review results, and the final assessment conclusion.

[0044] Example: Taking a mountain wind farm as an example, this wind farm comprises 32 5.0MW wind turbine units, with concrete tower structures. Each concrete tower has a prestressed concrete lower section and a steel upper section, with a steel-concrete transition section in between. The wind farm's roads are branched, with some turbine locations requiring more than 2km of road transport distance. Inspection equipment includes an industrial-grade drone platform, a rope-assisted wall-climbing robot, and a ground analysis platform.

[0045] like Figure 1As shown, the method in this embodiment includes the following steps.

[0046] S1. Construct a multimodal inspection dataset for wind power concrete towers.

[0047] Images were collected at the wind farm site of concrete tower sections, steel tower sections, and transition connection sections. Simultaneously, samples of varying widths of cracks, areas of spalling, coating damage, localized corrosion, and flange water seepage were prepared on laboratory test components. Data acquisition equipment included a UAV visible light camera, a UAV infrared thermal imager, a wall-climbing robot close-range camera, a wall-climbing robot infrared thermal imager, and a laser rangefinder. During data acquisition, the camera location number, relative tower base height, circumferential angle, acquisition time, ambient temperature, wind speed, and the acquisition equipment number were recorded.

[0048] For concrete tower section samples, vertical cracks, circumferential cracks, diagonal cracks, spalling, exposed reinforcement, and water seepage were labeled. For steel tower section samples, corrosion, coating damage, weld anomalies, and temperature anomalies were labeled. For transition connection section samples, abnormal bolt signs, flange misalignment, water seepage marks, and localized corrosion were labeled. Crack labels included the crack centerline, crack boundary, and maximum width; spalling labels included the spalling outline and maximum spalling depth; corrosion labels included the corrosion area outline and area. Image samples were stratified according to damage type, tower section type, and acquisition equipment, with a training set, validation set, and test set ratio of 7:2:1.

[0049] S2, train the end-side damage rapid screening model and the ground-side multimodal semantic analysis model and deploy them.

[0050] The rapid end-side damage screening model employs a lightweight convolutional backbone network, a feature fusion module, a target detection branch, a semantic segmentation branch, and an infrared anomaly discrimination branch. The UAV-based model takes a 640×640 resolution visible light image and infrared thermal image as input, and outputs the suspected damage bounding box, category, confidence score, and infrared temperature difference. The wall-climbing robot-based model takes a close-range high-resolution image, infrared thermal image, and laser ranging results as input, and outputs a pixel-level damage mask, crack width, spalling area, corrosion area, and connectivity misalignment.

[0051] The ground-based multimodal semantic analysis model employs a visual encoder, a structured numerical feature encoder, and a text generation module. Input data includes damage images, segmentation masks, damage geometric measurements, infrared temperature differences, unit location numbers, tower segment types, relative tower base height, ambient wind speed, ambient temperature, and historical inspection records. Output fields include damage name, damage area, measurement value, risk level, historical rate of change, recommended re-inspection cycle, and remedial measures. The edge model is deployed on the edge computing modules of UAVs and wall-climbing robots, while the ground-based multimodal semantic analysis model is deployed on the ground analysis platform workstation.

[0052] S3. Perform rapid cross-site screening of wind farms using drones.

[0053] Before the drone inspection, the planar coordinates of 32 turbine locations were read from the wind farm's overall plan, and a set of turbine locations to be inspected was generated by combining the take-off and landing point locations, tower heights, road and no-fly zone information. The drone path optimization target was constructed using formula (1), and the following settings were configured: , , When a single battery charge is insufficient to cover all the aircraft positions, the 32 aircraft positions are divided into 4 inspection batches, with each batch containing 8 aircraft positions.

[0054] The drones fly to their designated locations along a planned path, acquiring visible light and infrared thermal images within a range of 15m to 20m from the outer surface of the tower. The drone's edge computing module performs real-time inference on the acquired images, outputting suspected damage bounding boxes, categories, and confidence levels. For suspected damage with a confidence level greater than 0.6, the drone's location number, relative tower base height, circumferential angle, shooting pose, visible light thumbnail, and infrared temperature difference are recorded. After completing one batch, the drone uploads the screening results to the ground analysis platform.

[0055] S4. Calculate the damage risk score of a single tower structure and generate a review priority queue.

[0056] The ground analysis platform summarizes the UAV screening results and extracts the area ratio, estimated crack width, infrared temperature difference, vertical position, tower segment type, and historical change information of suspected damage at each location. Equation (2) is used to calculate the single-tower structure damage risk score. In this embodiment, the weight of a through-crack in concrete is 1.00, the weight of misalignment in the transition connection section is 0.95, the weight of abnormal welds in the steel tower section is 0.90, the weight of concrete spalling is 0.70, the weight of localized corrosion is 0.55, and the weight of coating damage is 0.40. The weight of the transition connection section tower segment is 1.30, the weight of the concrete tower segment is 1.10, and the weight of the steel tower segment is 1.00. The normalized value of the damage development rate is... It is calculated from the current measurement value and the historical inspection measurement value.

[0057] For example, a circumferential crack and a spalling were found at the WT08 site in the concrete tower section. The crack width was estimated at 3.0 mm, and the spalling area accounted for 0.018 of the local surface area. Historical records show that the crack width at the same location was 2.1 mm. The ground analysis platform calculated a structural damage risk score of 0.78 for WT08. After sorting the risk scores from high to low, a dynamic priority queue for the wall-climbing robot's review was generated.

[0058] S5. Generate a scheduling scheme for the wall-climbing robot to cross the tower for verification.

[0059] The ground analysis platform reads the current position of the climbing robot, the wind farm road distance matrix, the equipment endurance status, the single tower operation time constraint, and the review priority queue generated in step S4. The cross-tower review scheduling objective function is established using equation (3). For turbine positions with higher risk scores, the risk priority term in the objective function takes a smaller cost, so that high-risk turbine positions are given priority in the preceding review tasks.

[0060] In this embodiment, the wall-climbing robot is transported by a maintenance vehicle. The installation and dismantling time for a single tower is 30 minutes, and the verification operation time for a single tower is 45 to 90 minutes. The first-day verification sequence generated by the ground analysis platform is WT08, WT21, and WT15, and the suspected damage area, relative tower base height, circumferential angle, UAV image evidence, and estimated operation range for each position are sent to the wall-climbing robot control terminal.

[0061] S6, a rope-assisted wall-climbing robot, performs close-range verification and dimensional measurement.

[0062] After completing rope anchoring and adsorption-based walking unit checks at the designated inspection location, the wall-climbing robot begins its operation. It moves along the outer surface of the tower to near the suspected damage height and circumferential angle identified by the drone screening. The attitude adjustment unit brings the camera's optical axis as close as possible to the tower surface normal. For cracks in the concrete tower section, the robot acquires near-field visible light images and laser spatial measurement data, establishing a pixel scale conversion relationship based on camera intrinsic parameters, working distance, and pixel size. After segmenting and masking the cracks, the robot calculates the crack pixel width along the crack skeleton normal and converts it to the actual crack width.

[0063] At WT08, the wall-climbing robot verified that the maximum width of the circumferential crack was 3.2 mm, the crack length was 680 mm, the adjacent peeling area was 320 cm², and the maximum peeling depth was 18 mm. For the steel tower section at WT21, the robot identified three rusted areas, with the largest rusted area being approximately 150 cm², and the infrared temperature difference between the rusted area and the adjacent coating area was 1.8℃. For the transition connection section at WT21, the robot collected multi-angle images of the flange and bolt areas, measuring a flange water seepage mark approximately 15 mm wide and 120 mm long, and a misalignment of approximately 2 mm at one connection point.

[0064] S7. Generate a structured inspection assessment report and update the damage database.

[0065] The ground-based analysis platform inputs UAV screening data, wall-climbing robot verification data, basic information on the UAV positions, tower segment type information, and historical inspection records into a ground-based multimodal semantic analysis model. The model first performs a consistency check on the UAV screening category and the wall-climbing robot verification category; then, it determines the risk level based on the structural damage grading rules for concrete tower segments, steel tower segments, and transition connection segments; next, it compares the current measurement values ​​with historical inspection records to calculate the damage change rate; finally, it generates a structured inspection assessment report.

[0066] Taking WT08 as an example, the ground analysis platform outputs the following structured record: the unit number is WT08; the tower section is a concrete tower section, approximately 45.2m from the tower base; the damage type is circumferential crack and local spalling; the maximum crack width is 3.2mm, and the crack length is 680mm; the spalling area is 320cm², and the spalling depth is 18mm; compared with the record from 6 months ago, the crack width has increased from 2.1mm to 3.2mm; the risk level is high; the recommended maintenance priority is Level 1 maintenance; the recommended treatment measures are crack sealing grouting, spalling area repair, and shortening the subsequent re-inspection cycle.

[0067] The ground analysis platform also generates a damage distribution map in the form of a unfolded diagram of the concrete tower, marking the damaged areas, damage types, and risk levels on the unfolded diagram. Simultaneously, it generates a review task record form, recording the consistency between the UAV screening results, the wall-climbing robot review results, and the final assessment conclusion. The structured inspection assessment report and damage distribution map are written into the wind farm's concrete tower damage database, serving as the data source for calculating the normalized damage development rate value in the next inspection cycle.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

Claims

1. A method for optimizing inspection and assessing damage in hybrid tower wind farms, characterized in that, Includes the following steps: Drones were used to screen and photograph the towers of wind farms, and to collect screening data of the towers. The screening data of the tower is input into the edge computing module of the UAV for rapid screening of end-side damage, and the suspected damage area of ​​the tower, damage category, confidence level and the corresponding UAV position number are output. Based on the suspected damage area, damage type, confidence level, and the corresponding drone position number, computer vision algorithms are used to analyze the area ratio of the suspected damage area of ​​the tower, crack width, infrared temperature difference, spatial location of the suspected damage, tower segment type, and structural sensitivity indicators based on the images taken by the drone. Based on the finite element numerical model of a single tower, the stress and dynamic response indices of each tower structure are obtained through simulation calculation. Based on the degree of structural damage simulated, the structural damage risk score of each tower is calculated, and a priority queue for review of the climbing robot is generated according to the structural damage risk score. Based on the review priority queue, combined with the current position of the wall-climbing robot, the road distance matrix of the wind farm, the endurance status of the wall-climbing robot and the single tower operation time constraint, a cross-tower review scheduling scheme for the wall-climbing robot is generated. The wall-climbing robot reaches the tower with suspected damage according to the cross-tower verification scheduling scheme, performs close-range verification of the suspected damage area of ​​the tower, and collects verification data of the suspected damage area; The ground analysis platform inputs the screening data, verification data, aircraft position information, tower segment type and UAV historical inspection records into the ground-based multimodal semantic analysis model, and generates a structured inspection assessment report by combining the structural damage grading rules.

2. The inspection optimization and damage assessment method for hybrid tower wind farms as described in claim 1, characterized in that, The rapid end-side injury screening model and the ground-based multimodal semantic analysis model were constructed as follows: Visible light images, infrared thermal images, and spatial measurement data were collected from the concrete tower section, steel tower section, and transition connection section of the tower to construct a multimodal inspection dataset containing damage category, spatial location, and damage scale annotations. A rapid end-side damage screening model and a ground-side multimodal semantic analysis model were trained based on a multimodal inspection dataset. The rapid end-side damage screening model is deployed on the edge computing modules of UAVs and rope-assisted wall-climbing robots, while the ground-based multimodal semantic analysis model is deployed on the ground analysis platform. The rapid end-side damage screening model outputs the suspected damage area, category, and basic scale in real time, while the ground-based multimodal semantic analysis model is used for multi-source inspection data fusion analysis and structured damage assessment.

3. The inspection optimization and damage assessment method for hybrid tower wind farms as described in claim 2, characterized in that, The specific steps for constructing a multimodal inspection dataset are as follows: Visible light images, infrared thermographs, and spatial measurement data were collected from concrete tower sections, steel tower sections, and transition connection sections at wind farm sites or in the laboratory tower test components. Based on the collected visible light images, infrared thermal images and spatial measurement data, vertical cracks, circumferential cracks, diagonal cracks, spalling, exposed reinforcement and water seepage in the concrete tower section are marked; rust, coating damage, weld abnormalities and temperature abnormalities in the steel tower section are marked; and loose bolts, flange misalignment, water seepage marks and local rust in the transition connection section are marked. Establish a correlation record for each damage on the tower, including damage type, tower segment type, relative tower base height, circumferential angle, geometric dimensions, acquisition equipment, and acquisition time; The multimodal inspection dataset obtained after establishing the association records is organized and divided into layers according to damage type, tower segment type, and acquisition equipment.

4. The inspection optimization and damage assessment method for hybrid tower wind farms as described in claim 3, characterized in that, The rapid end-side damage screening model includes a target detection branch, a semantic segmentation branch, and an infrared anomaly discrimination branch. The target detection branch is used to output the bounding box and category of the damage. The semantic segmentation branch is used to output the pixel-level mask of crack, peeling, and corrosion areas. The infrared anomaly discrimination branch is used to determine whether the temperature is abnormal based on the infrared temperature difference between the damaged area and the adjacent normal area.

5. The inspection optimization and damage assessment method for hybrid tower wind farms as described in claim 1, characterized in that, Before screening and photographing the towers of a wind farm, the drone pre-generates an inspection path. The drone then performs screening and photographing along the inspection path. The inspection path is generated based on the spatial coordinates of each drone position within the wind farm, the tower height, the drone's endurance constraints, the take-off and landing point locations, and the wind farm's meteorological conditions. The screening data collected by the drone includes visible light images, infrared thermal images, and shooting pose data of the outer surface of the tower.

6. The inspection optimization and damage assessment method for hybrid tower wind farms as described in claim 1, characterized in that, The method for collecting crack width data in the verification data is as follows: The working distance from the camera of the wall-climbing robot to the surface of the tower is determined based on laser spatial measurement data; Establish pixel scale transformation relationship by combining camera intrinsic parameters and image pixel size; The semantic segmentation mask of the crack is skeletonized, the crack pixel width is calculated along the skeleton normal, and the crack pixel width is converted into the actual crack width.

7. The inspection optimization and damage assessment method for hybrid tower wind farms as described in claim 1, characterized in that, The ground-based multimodal semantic analysis model takes damage images, segmentation masks, damage geometric measurements, infrared temperature differences, station numbers, tower segment types, relative height of the damaged area to the tower base, ambient wind speed, ambient temperature, and historical inspection records as inputs. It outputs an inspection assessment report in the form of structured fields. The structured fields include at least the damage name, damage area, measurement value, risk level, historical change rate, recommended re-inspection cycle, and treatment measures.

8. The inspection optimization and damage assessment method for hybrid tower wind farms as described in claim 1, characterized in that, The structural damage grading rules are classified according to different tower section types, including: For concrete tower sections, they are classified according to crack width, crack length, crack direction, spalling area, spalling depth, and water seepage status; For steel tower sections, classification is based on rust area, coating damage area, abnormal weld location, and degree of temperature abnormality; For transitional connection sections, classification is based on the amount of misalignment at the connection points, signs of bolt abnormalities, water seepage marks on the flanges, and localized corrosion.

9. The inspection optimization and damage assessment method for hybrid tower wind farms as described in claim 1, characterized in that, The screening data from the drone and the verification data from the wall-climbing robot both contain unified spatial identification information, which includes the drone position number, tower segment type, relative tower base height, circumferential angle, and collection time, in order to match the records of the same damage in the drone and the wall-climbing robot, as well as in different inspection cycles of the two.

10. The inspection optimization and damage assessment method for hybrid tower wind farms as described in claim 1, characterized in that, When generating a structured inspection and assessment report, the ground analysis platform also generates a damage distribution map and a review task record table. The damage distribution map uses the tower unfolded diagram as a base map to indicate the location, type, and risk level of each damage. The review task record table records the consistency between the UAV screening results, the wall-climbing robot review results, and the final assessment conclusion.

Images