Unmanned Aerial Vehicle (UAV)-Assisted Damage Detection of Wind Turbine Blades: A Review

Abstract

The wind energy sector is experiencing rapid growth, marked by the expansion of wind farms and the development of large-scale turbines. However, conventional manual methods for wind turbine operations and maintenance are struggling to keep pace with this development, encountering challenges related to quality, efficiency, and safety. In response, unmanned aerial vehicles (UAVs) have emerged as a promising technology offering capabilities to effectively and economically perform these tasks. This paper provides a review of state-of-the-art research and applications of UAVs in wind turbine blade damage detection, operations, and maintenance. It encompasses various topics, such as optical and thermal UAV image-based inspections, integration with robots or embedded systems for damage detection, and the design of autonomous UAV flight planning. By synthesizing existing knowledge and identifying key areas for future research, this review aims to contribute insights for advancing the digitalization and intelligence of wind energy operations.

1. Introduction

In recent years, there has been a growing emphasis on optimizing energy structures and reducing greenhouse gas emissions. Consequently, the development of renewable energy sources, particularly wind energy, has gained significant momentum. According to the Global Wind Energy Council (GWEC), the global installed capacity of wind power surged from over 300 GW to 906 GW by 2022, reflecting a remarkable cumulative growth of nearly 200%, as shown in Figure 1 [1]. While this rapid expansion has significantly advanced the wind power industry, it has also introduced several challenges that must be effectively addressed.

Wind turbines are essential components for harnessing wind energy and converting kinetic energy from the wind into mechanical power to generate electricity. It is well-noted that the development of wind turbines has advanced significantly over the past few decades, driven by economic benefits and environmental concerns [2]. Modern wind turbines have become larger and more efficient, with substantial increases in rotor diameter, hub height, and overall capacity [3]. The expansion of wind farms, both onshore and offshore, has been pivotal in scaling up wind energy production. The trend toward larger wind turbines aims primarily at capturing more energy by accessing higher and more consistent wind speeds. As demonstrated in Figure 2, the size of wind turbines has increased significantly over the years, with rotor diameters exceeding 150 m and hub heights reaching up to 200 m becoming common [4]. While these larger turbines offer substantial benefits in terms of energy yield, they also present significant challenges concerning operation and maintenance (O&M). One primary difficulty is the accessibility of turbine components for maintenance. Larger turbines require more complex and higher-capacity lifting equipment for installation and repairs, thus increasing logistical complexity and costs [5]. Moreover, the sophistication of larger turbines, equipped with advanced control systems and sensors, necessitates specialized knowledge for troubleshooting and repair. Ensuring maintenance personnel are adequately trained to handle these systems is essential but can be resource-intensive [6].

The expansion of wind farms, particularly offshore, introduces additional O&M challenges. One significant challenge is the increased difficulty and cost of accessing offshore turbines for routine maintenance and emergency repairs. Adverse weather conditions and rough seas can limit access to offshore wind farms, potentially leading to longer downtime and reduced energy production [10]. This necessitates developing robust predictive maintenance strategies and remote monitoring technologies to minimize the need for physical interventions [11]. Economically, the increased O&M costs associated with larger turbines and expanded wind farms can affect the overall cost-effectiveness of wind energy projects. Technically, the challenges posed by larger turbines and offshore wind farms drive innovation in O&M practices, such as advances in remote sensing, autonomous drones for inspection, and predictive maintenance algorithms.

Wind turbine blades are the components with the highest incidence of failure in wind turbines. According to damage incident statistics, blade failures accounted for approximately 19.4% of all damage incidents worldwide as of 2012, as shown in Figure 3 [12]. On the other hand, wind turbine blades are among the most expensive and valuable components, constituting 15–20 percent of the total manufacturing cost of a wind turbine [13]. To prevent catastrophic failures and ensure the blades operate in optimal condition, it is crucial to implement effective blade damage detection systems.

Effective operation and maintenance (O&M) strategies are essential for ensuring the reliability, efficiency, and longevity of wind turbines. Traditionally, manual inspection methods have mostly been employed. While it is conventional, it presents several disadvantages. These inspections are labor-intensive, time-consuming, and costly, requiring skilled personnel, which increases operational expenses [15]. As wind turbines continue to increase in size, particularly those in offshore installations, reaching certain parts of the turbine becomes increasingly challenging and expensive. This can delay inspections and maintenance activities, potentially leading to undetected faults [16]. Furthermore, the quality and thoroughness of manual inspections can vary based on the technician’s expertise and environmental conditions, leading to missed defects and unreliable maintenance decisions [17]. In contrast, the advent of advanced technologies, such as autonomous drones, has significantly enhanced O&M practices. The use of unmanned aerial vehicles (UAVs), especially multi-rotor UAVs, offers several distinct advantages in terms of wind turbine inspection. UAVs can conduct inspections more quickly and cost-effectively, reducing the need for extensive human labor and mitigating the risks associated with accessing hard-to-reach areas. They provide consistent and high-quality data unaffected by the variability in technician expertise or environmental conditions, thus ensuring more reliable maintenance decisions and enhancing the overall efficiency of wind turbine O&M practices.

The primary focus of this paper is to reflect and summarize state-of-the-art achievements in the application of UAVs in damage detection of wind turbine blades, with particular emphasis on their usability in visual inspection. The structure of this paper is illustrated in Figure 4.

2. Damage of Wind Turbine Blades

Wind turbine blades are composed of composite materials, including primarily glass-fiber-reinforced resin for the blade housing, carbon fiber for the blade tip and girder, and sandwich structure composite material for the leading and trailing edges, to meet the requirements of efficiency and strength-to-weight ratio [18,19,20]. Wind turbine blades are critical components responsible for capturing kinetic energy from the wind and converting it into mechanical energy [21,22]. These blades are subjected to dynamic loading, extreme environmental factors, and material fatigue, which can lead to various forms of damage, such as cracks, debonding, and delamination [12,23,24]. Damage to wind turbine blades caused by extreme weather events, primarily lightning, storms, and strong winds, accounts for over 76% of total damages, excluding other unknown causes [12]. Such damage increases the likelihood of blade failure, thereby elevating safety risks, reducing blade lifespan, and potentially leading to the collapse of the entire structure [21,25,26]. The resulting decrease in aerodynamic performance reduces power generation [27,28,29,30], while increased surface roughness due to damage elevates noise levels [31,32].

Typical damages to wind turbine blades and their causes and impacts include debonding, cracking, contamination, erosion/corrosion, delamination, and splitting. These damages often result from a combination of factors, such as material fatigue, environmental influences, manufacturing defects, and human error. This section focuses on the primary causes and likely mechanisms of these damages.

Wind turbine blades consist of skin and webs, which are typically assembled using bonding methods. Debonding, the loss of adhesion between materials [33], is a common type of blade damage and often the initial failure mechanism that can lead to progressive blade collapse [34]. Depending on the location, debonding can occur in the skin and main spar flanges, between the upwind and downwind skins, between the skins and the core, or on the skin surface. These can be categorized as skin/adhesive debonding, adhesive joint failure, sandwich debonding, and surface coating debonding [35]. Structural cracks often start in fatigue-critical areas due to cyclic gravitational bending, shear, axial loading, and centrifugal forces [36,37]. Cracks not only increase surface roughness but, upon formation and growth, can lead to catastrophic failure due to material fatigue degradation [37,38]. Surface contamination from dust, grease, insects, and other factors increases blade roughness [39,40] and promotes a corrosive environment. Material wear damage can be divided into erosion (physical wear) and corrosion (chemical wear). Both types alter the aerodynamic shape of the blade and increase roughness, thereby reducing aerodynamic efficiency [41,42]. Moreover, the leading edge of the blade, particularly near the tip where velocities are highest, is highly susceptible to erosion and can even experience delamination [43,44]. This erosion not only degrades the material but can also propagate delamination along the affected area.

Delamination in wind turbine blades occurs due to failures in bonding between composite layers, a process prone to errors due to the high manual component of current manufacturing practices [45]. Splitting and fracturing of individual fibers in the laminate layers of the skin and the main spar fiber result from insufficient tensile and compressive strength. Delamination and splitting, as internal damages, cause stress concentration in areas with discontinuities, thereby reducing blade stiffness and load-carrying capacity [31,46]. Note that wind turbine blades are particularly vulnerable to lightning strikes, making lightning protection systems essential for mitigating damage. However, damage detection in the regions of lightning attraction and peripheries is crucial to identify failures in lightning receivers, as damages and cracks often develop in these areas [47,48]. Vortex generators, typically fixed passively on wind turbine blades, are subject to cyclic stress-induced fatigue, which can result in panels with missing teeth, thereby reducing the aerodynamic performance of the blade.

Typical damages of wind turbine blades are listed in

Table 1. Damage types for wind turbine blades [25,35,37,39,43,46].

Damage Type	Details of Typical Damage Types	Internal/Outer
Debonding 1	Skin/adhesive or main spar/adhesive layer debonding	Outer
Debonding 2	Adhesive joint failure between skins	Outer
Debonding 3	Sandwich panel face/core debonding	Outer
Debonding 6	Skin/adhesive debonding induced by buckling	Outer
Debonding	Coating debonding, such as gelcoat off	Outer
Crack	Crack in split crack 5, gelcoat 7, early crack	Outer
Contamination	Dirt, oil dirt, insect contamination	Outer
Erosion/corrosion	Edge erosion, leading edge erosion/rust, pitting	Outer
Delamination	Delamination driven by a tensional or a buckling load 4	Internal/outer
Splitting	Fiber failure in tension, laminate failure in compression 5	Internal/outer

Annotation: the superscript represents the corresponding damage type in Figure 5.

, with a schematic illustration in Figure 5 and real images in Figure 6. Due to the structural characteristics of the blades, different types of damage are distributed across various regions. The most vulnerable locations include the root, the 30–35% spanwise position, the 70% spanwise position, and the adhesive or welded joints [7,33]. These areas are particularly susceptible to damage due to their critical structural roles and the stresses they endure.

3. Blade Damage Detection Methods

Wind turbines are often subject to various forms of damage due to their exposure to harsh environmental conditions and continuous operational stresses. Effective damage detection is essential for maintaining structural integrity, minimizing downtime, extending turbine lifespan, improving safety, and optimizing performance. Early detection of such damage is critical to ensuring the structural integrity and operational safety of wind turbines [15]. Traditional blade damage detection methods primarily include visual inspection, thermography, strain measurement, vibration analysis, ultrasonic testing, and acoustic emission monitoring.

Table 2. Various damage detection technologies for wind turbine blades with their homologous types of sensors, data indicators, arrangement forms, and precision [14,25,53,54,55].

Technology	Sensors	Indicators	Arrangement	Precision
Vision	Binoculars, eyes, professional tools	/	On ground/hanging basket/climber on wind turbine	/
	Camera	Image feature	On ground	≈1 mm
			UAV	/
Thermography	Thermal/infrared camera	Temperature Thermal energy Thermal responses	On ground	3–5 mm
			UAV	/
Strain measurement	Strain gauges, FBGs	Strain Peak strain Strain rate Deflection	On wind turbine	≈1 cm
	Camera		UAV	/
Vibration	Displacement sensors/velocity sensors/accelerometers	Frequency–response Amplitude Mode shape Strain energy Spectral kurtosis	On wind turbine	<1 mm
	Radar		On ground	<1 mm
	Camera		UAV	/
Ultrasound	Piezoceramic material sensors	Frequency Amplitude Time-of-flight Reflection energy	On wind turbine	≈1 cm
	Robots		UAV	/
Acoustic emission	AE sensors/macro-fiber composite sensors/piezoceramic material sensors	Acoustic emission signals Waveform characteristics Acoustic energy RMS	On wind turbine	<1 cm
	Embedded acoustic emission sensors		UAV	/

Annotation: “On wind turbine” means that sensors rely on surface contact with wind turbines or are embedded in wind turbine blades, usually with high accuracy and reliability through continuously direct measurement. Image-based damage detection can usually reach the pixel level, which is determined by the camera resolution, measurement settings, environmental conditions, image processing algorithms, etc. UAV-based detection studies are still at an early stage and prone to feasibility testing through good relevance and consistency with other technology, while there is a negligible focus on their use for the direct description of the size of the damage and the precision of the technique.

summarizes the sensor types, data indicators, arrangement forms, and precision levels related to these detection technologies.

Table 3. Comparison and discussion of various technologies in damage detection of wind turbine blades [14,15,25,49,56].

Technology	Advantages	Limitations	Damage Types	Principle
Vision	Low cost; independence from system complexity; visuality	Heavy computation; poor explanation; susceptible to environment	Surface damage: crack, debonding, erosion, etc.	Based on real or similar human vision
Thermography	Full field measurement; visuality; fast detection	Unable to detect early damages; susceptible to environment	Fatigue, delamination, etc.	Focus on thermodynamic property variations
Strain measurement	Continuous monitoring; external-incentive-free; wide sampling rates	Accuracy subjected to selected area; one sensor for one point; low fatigue resistance	Minute changes in length or deformation	Strain gauges: resistance change detection according to grid variation
Vibration	High sensitivity; easy deployment	Unable to detect early damages; susceptible to environmental disturbances	Frequency response, modal parameters, etc.	Focus on vibration representing dynamic properties
Ultrasound	Available for location, size, and depth of inner damage	Complicated data processing; time-consuming	Inner structural damage: Delamination, debonding, etc.	Wave reflection detection according to differences in the material and its damage
Acoustic emission	Continuous monitoring; high sampling frequencies; high sensitivity	High cost; high sensor deployment requirement; complicated data processing	Damage initiation, crack propagation, plastic deformation	Transient elastic wave detection through rapid release of energy from local sources

Annotation: NDT represents non-destructive testing. “Principle” column of the table neglects the specific electrical signal transformation. The descriptions in the table do not include UAV methods.

presents a summary of the advantages, limitations, detected damage types, and principles associated with the aforementioned detection technologies.

3.1. Vision

Most of the damage to wind turbine blades occurs on the exterior and can be visually observed. Due to its high deployment efficiency and low cost, visual inspection is a common approach for detecting damage in wind turbine blades [57]. Manual inspection involves professionals using the naked eye or telescopic methods to observe and diagnose blade damage either from the ground, from hanging baskets, or by climbing on the turbines [58]. Particularly for large-scale wind turbines, this method requires highly skilled personnel and involves significant safety risks. In the meantime, detection efficiency is subject to personnel fatigue and proficiency.

Comparatively, with the rapid advancements in camera technology and computer vision, machine vision inspection using high-definition images has become a promising alternative to manual visual inspection. This approach significantly reduces the need for manual intervention in hazardous environments [59]. Machine vision employs computer science and image processing algorithms to simulate the human stereoscopic vision system, enabling automatic scene detection, analysis, and interpretation [60]. Damage detection using machine vision can be conducted on single-view images, which are suitable only for preliminary damage identification [61]. To obtain comprehensive information about the location, size, and shape of the damage, multi-perspective and multi-position views are necessary, especially for wind turbine blades, which are large structures with complex shapes [62]. In addition to detecting surface damage, such as cracks and debonding, machine-vision-based methods are also applicable to strain and vibration measurement. Techniques including digital image correlation (DIC) and optical flow can be used with image sequences to track targets and obtain the dynamic characteristics of wind turbine blades [63]. However, there are two critical challenges that need to be addressed: image quality and processing efficiency.

Due to the contradiction between the remote shooting distance and the requirement for high resolution in ground-based cameras, UAV technology has emerged as an effective means of visual inspection. UAVs offer multi-view and high-definition imaging capabilities due to their high mobility. Image processing is crucial for the accuracy and effectiveness of wind turbine blade operations and maintenance [64]. Consequently, image processing algorithms have become a primary focus for researchers, including certain areas, such as deep learning [65,66], motion deblurring [67], and adaptation to complex backgrounds and changing environments [68].

3.2. Thermography

Thermography-based detection technology focuses on identifying variations in the thermodynamic properties of wind turbine blades [69]. Material damage on the blade typically causes local temperature anomalies, which can be detected non-destructively by measuring temperature gradients [70]. This method is effective for visualizing and scanning the full-field surfaces of large wind turbine blades using infrared devices. However, it is less suitable for early damage detection due to the slow development of temperature changes [71]. The accuracy of thermographic detection relies heavily on thermal image processing, where it is crucial to emphasize the impact of blade damage on temperature while eliminating interference from other factors.

3.3. Strain Measurement

Strain-measurement-based detection technology focuses on detecting minute changes in the length or deformation of wind turbine blades by using strain sensors under external excitation [56]. Direct strain (ε = x/l) and shear strain (γ = x/d) are utilized in the inspection of wind turbine blades [72]. Strain sensors, often installed on the surface or embedded within the layers of the blade, allow for indirect and continuous monitoring of structural damages through the expansion or contraction of the blade over the long term [73]. However, the accuracy and sensitivity of strain measurements are highly dependent on the deployment strategy, particularly the distance between the sensor location and the damage [25].

Strain gauges detect damage through resistance changes caused by grid variation, whereas fiber Bragg gratings (FBGs) measure the variation in the length of the fiber grating [74]. FBGs, which can be embedded in composites, offer advantages, such as small size and high performance. Despite these benefits, strain gauges remain more widely used due to their lower cost, even though they lack the noticeable advantages of FBGs [75]. Strain gauges must be mounted on the blade surface and are prone to failure under long-term operation [76]. Therefore, the development of low-cost, high-accuracy, and wireless sensors, along with advanced pattern recognition techniques, is imperative.

3.4. Vibration

Vibration-based detection technology is a pivotal method for assessing the structural health of wind turbine blades by monitoring abnormal vibrations indicative of damage or irregularities [54]. This technique is based on the analysis of dynamic properties, such as frequency–response and modal parameters, which are altered by the presence of damage. Despite its complexity and traditional limitations in real-time application, advancements in intelligent data processing have enabled continuous and online damage detection, thus enhancing the reliability of wind turbine blade operations [77]. Sensors, primarily displacement sensors, velocity sensors, and accelerometers, are strategically placed on blade surfaces to capture vibration signals across a spectrum of frequencies [78] Meanwhile, the integration of ground-based radar (GBR) as a non-contact sensor offers a promising approach for remote vibration detection suitable for large-scale and offshore blade inspections [53].

The application of advanced analytical techniques for frequency, time, and time–frequency analyses, such as wavelet transform, is essential for extracting unique signatures that facilitate damage identification [79,80]. The technology’s advancement hinges on the development of sophisticated and interpretable algorithms to overcome environmental interferences and improve detection immediacy. The incorporation of non-contact sensors enhances remote monitoring capabilities [80,81,82].

3.5. Ultrasound

Ultrasound-based detection is a prevalent non-destructive testing (NDT) technique that leverages the propagation and reflection of elastic waves within materials to identify internal structural damage, such as delamination and debonding [7,78]. This method relies on detecting reflected waves from damage sites, which are transmitted through the material and captured on opposite surfaces [45]. Analysis of reflection, attenuation, resonance, and transmission patterns provides insights into the size, location, and properties of the damage [71]. Ultrasonic Guided Waves (UGWs) are particularly effective for pinpointing damage with precision at the millimeter level [83].

Innovations in this field include non-contact laser ultrasonic scanning and automated pulse-echo ultrasound systems, which enhance inspection capabilities for wind turbine blades. Applications extend beyond damage detection to include de-icing systems for wind turbine blades, where UGWs are combined with low-frequency vibrations for ice removal [84,85]. Despite their precision, fast scanning speeds, and effective flaw detection capabilities, challenges remain, such as the complexity of signal processing, the extended time required for data acquisition, and the need for surface contact [33,86,87].

3.6. Acoustic Emission

Acoustic emission (AE) detection employs sensors, such as macro-fiber composite sensors or piezoceramic material sensors, to capture high-frequency transient elastic waves generated by the release of energy from micro-structural damage within wind turbine blades [23,88]. These sensors convert the waves into electrical signals that can be analyzed to estimate the criticality of damage, including initiation, propagation, and failure processes. This method is advantageous for its rapid, efficient, and non-invasive nature, providing earlier damage detection compared to vibration-based techniques [78,89].

AE technology is increasingly popular for monitoring material and structural health, effectively tracking damage expansion and assessing the timing, location, and severity of damage during wind turbine operation. However, the technique requires a substantial number of sensors to be strategically placed on the wind turbine blades for optimal detection near potential damage sites [23,90]. The lack of a direct physical link between the AE signal and the specific damage, along with the need for high-sampling-frequency data acquisition systems, complicates signal processing and increases costs [33,78]. Furthermore, AE technology does not provide insights into the internal structural stress of the blades [72].

Machine vision and thermography technologies enable the visualization of blade damage with easy deployment and short-term detection periods. The accuracy of these methods hinges on the quality of the images and the effectiveness of image processing techniques. Conversely, strain measurement, vibration, ultrasound, and acoustic emission technologies are suitable for continuous monitoring, requiring the strategic deployment of sensors on or within the wind turbine blade. The accuracy and efficiency of these detection methods depend on thorough data cleansing to highlight damage-induced information while eliminating noise interference. Ultrasound and acoustic-emission-based detection are more direct for identifying internal damage, whereas thermography and vibration analysis provide circumstantial evidence that must be interpreted within the context of potential internal defects. Strain-measurement-based detection can indicate changes resulting from damage but is less capable of directly localizing damage compared to other methods. Visual inspection is primarily effective for detecting surface blade damages.

Currently, the direct placement of sensors in wind turbines is considered one of the most reliable and accurate methods for condition monitoring. The effectiveness of a condition monitoring system is inherently linked to the number and type of sensors employed [91]. However, challenges, such as high costs, deployment difficulties, and sensor faults in harsh environments, must be addressed. The Supervisory Control and Data Acquisition (SCADA) system is a crucial component of modern wind turbines, enabling continuous monitoring through data mining from up to 150 types of operating indicators. Despite this, the installation of sensors on wind turbines is associated with a high failure rate. To ensure the reliable functioning of wind turbines and their sensor systems, optimizing electromagnetic compatibility measures is imperative. Fontanes et al. conducted a comprehensive study on blade tip measurements in a laboratory setting, simulating blade conditions using UAV deployment with a vertical wire [92]. This investigation focused on understanding the atmospheric electric effects during fair weather, highlighting the polarization phenomenon observed between the blade and the wire. Based on their findings, an innovative artificial charge control system was proposed that can significantly reduce the potential of the vertical wire deployed by the UAV from 5 kV to 0 kV. They further analyzed the electrical effects of various blade tip geometries at different heights and vertical speeds, validating the performance of the UAV-deployed wire platform. Additionally, a model incorporating point/corona discharge and motion-induced current in wind turbines was formulated, enabling the simulation of the current under thunderstorm conditions [93].

Blade damage detection techniques primarily rely on vision, thermography, strain measurement, vibration, ultrasound, and acoustic emission by employing devices embedded in blades, mounted on the wind turbine, or deployed on the ground. Despite the potential and promising applications of these technologies, no single technique is universally optimal. Therefore, integrating multiple technologies tends to provide a more comprehensive damage detection approach, with the UAV-based methodology serving as a potential means. The framework for UAV-based damage detection of wind turbine blades, illustrated in Figure 7, encompasses the workflow of preflight preparation, flight operation, and postflight data processing. In addition, UAV inspection includes the take-off and landing positions, a schematic diagram of the UAV path, and the control mode. The subsequent section will detail specific research and development advancements of UAVs in optical photography, thermodynamic photography, and ultrasonic detection robot coupling.

4. Blade Damage Detection Using UAVs

Traditional methods of blade damage detection, though effective, present challenges related to safety, cost, and efficiency. The advent of UAVs offers a promising alternative with significant advantages over conventional methods. UAVs can carry various sensors and robots, facilitating comprehensive blade damage detection. Surface damage can be identified through visual inspections, while structural or internal damages can be monitored using thermography or ultrasound techniques. This section focuses on the application of UAVs in wind turbine blade damage detection and condition monitoring.

4.1. Optical-Image-Based Detection of Blade Damage

Aerial photography represents the most prevalent application of UAVs, supported by well-established technology. In the wind energy sector, intelligent image processing through machine vision is crucial for the detection of wind turbine blade damage. The following section outlines the fundamental workflow of image processing and subsequently reviews various types of blade damage and their associated research methods, including those for icing, strain and vibration, cracks, and erosion.

4.1.1. Elementary Image Processing Workflow

The utilization of UAVs equipped with optical imaging systems represents a groundbreaking approach in wind turbine blade damage detection. This approach capitalizes on the mobility, flexibility, and advanced imaging capabilities of UAVs to conduct detailed and comprehensive assessments of blade conditions. A critical factor in optimizing the optical-image-based detection of blade surface damage is the application of precise image-processing techniques. These techniques encompass a wide array of methods, including image filtering, enhancement, segmentation, detection, and classification.

Image filtering plays a crucial role in smoothing raw images, removing noise, and highlighting essential features. Among the various filters available, the adaptive median filter algorithm [94] is widely adopted for these purposes. Techniques, such as image deblurring, are employed to enhance image resolution and clarity. One effective method is super-resolution reconstruction using convolutional neural networks (CNNs) [95], which can notably enhance the visual quality of blurred images. Peng et al. conducted extensive research in this field, creating three datasets comprising pairs of sharp and blurred images of wind turbine models captured under varying rotational speeds and backgrounds [96]. The selection of images involved meticulous segmentation to minimize differences between sharp and blurred frames. The authors synthesized image pairs from video frames and simulated motion to assess the collective performance of DeblurGANv2 and Inception-ResNet-v2 models in deblurring and subsequent analysis.

Image segmentation involves dividing captured imagery into distinct feature or background areas tailored to specific scenarios and requirements for subsequent target detection. In the context of detecting minute features on wind turbine blades, images are often segmented into smaller patches. Several segmentation methods are utilized, including threshold segmentation [96], which categorizes pixels into different classes based on predefined thresholds; edge detection [97], which identifies boundaries between regions in an image; and semantic segmentation [98], which assigns semantic labels to pixels or regions. To enhance the detection of small damage on wind turbine blades, Gohar et al. introduced a slice-aided defect detection approach that utilizes segmentation based on slices rather than traditional patch-based segmentation [99]. Furthermore, an advanced blade image stitching technique has been proposed to improve edge segmentation and aid in defect deduplication [51]. This method employs a full-blade sequential segment mask, shape loss calculation, and shifting operations to effectively stitch together blade images, thereby enhancing the edge segmentation process. Despite challenges, such as fragile key points, textural features of the blade, and the unstable pose of the UAV, the use of the blade surface and shape features in image stitching provides superior visualization of the blade, as illustrated in Figure 8. Although there are limited studies on blade stitching in optical and thermal images, the results shown in Figure 8 demonstrate the potential for comprehensive full-blade damage detection.

Target detection involves locating, identifying, and classifying objects within an image. Wang et al. described the detection of crack regions using both original and extended Haar-like features [100] and proposed an extended cascading classifier for localization, coupled with the parallel Jaya K-means algorithm for boundary detection [101]. However, traditional computer vision techniques face limitations in recognizing multiple damage types and heavily rely on manual annotation. The advent of artificial intelligence and the evolution of machine learning have facilitated the adoption of these techniques for feature extraction and classification [102]. The transition to deep learning has further enhanced these capabilities. Figure 9 illustrates the results of detecting four types of defects using deep learning methods based on optical photography. Additionally, other defects, such as sand holes and vortex generator panels with missing teeth, can also be identified and classified, depending on specific datasets, classification criteria, and labeling standards.

4.1.2. Frozen Blade Detection and Ice Mitigation

Due to the high wind generation potential in cold climates, wind turbines frequently encounter icing events that significantly impact their performance. Even under conditions of high wind speed and a blade pitch angle of 90° (the angle between the blade and the rotating plane), ice accumulation can reduce the aerodynamic efficiency of the blades. This results in insufficient torque generation, affecting the turbine’s normal operation and causing up to 80% power loss [103]. On the other hand, the uneven load distribution due to icing can shorten the lifespan of the wind turbine, and the risk of ice fragments detaching from the blades poses safety hazards. Existing solutions for addressing this issue can be broadly categorized into three types: ice detection, anti-icing, and de-icing techniques.

Specifically, ice detection methods, such as mass change measurements and icing sensors, aim to provide early warnings of icing conditions. Anti-icing techniques, including the application of black paint or specialized surface coatings, are designed to prevent ice accumulation. De-icing techniques, ranging from manual inspection using rope access to the use of resistive heaters, focus on removing ice that has already formed on the turbine blades. The types of ice mitigation methods, their effectiveness, cost, increase in roughness, external energy consumption, and various stages of development are compared in

Table 4. Comparison of various methodologies for ice mitigation (anti-icing and deicing) of wind turbine blades [104,105,106,107].

Methodology	Type	Effectiveness	Cost	Roughness Increase	External Energy	Stage
Black paint	Anti-icing	Limited	Low	Low	None	Prototype
Coating	Anti-icing	Limited	Low	Medium	None	Prototype
Microwave/infrared/ultrasonic	De-icing	Effective	Medium	Almost none	Low	Experimental
Pneumatic/expulsive	De-icing	Effective	High	High	Low	Operational (aeronautics)
Chemicals	Both	Momentaneous and degrading	Low	Medium	Low	Experimental
Hot air	Both	Effective	High	None	High	Operational
Inside resistive heaters	Both	Effective	Medium–high	None	Low–medium	Experimental

Annotation: Experimental UAV-based deicing technology equipped with a sprayer and tethered cables (for long-term operation) belongs to the mobile spraying chemical methodology. The table encompasses the majority of anti-icing and de-icing methodologies, whereas some methodologies, such as active pitch control or electromagnetic pulse, are not included in the table for comparison and discussion.

.

The Norwegian University of Science and Technology (NTNU) has initiated research exploring the parallels between drones (UAVs) and wind turbines, particularly their applications in addressing wind turbine icing [108]. Gao et al. utilized advanced image processing techniques on high-resolution aerial drone imagery (approximately 3.5 mm/pixel) to accurately delineate the outer contours of ice accretion on the blades of utility-scale wind turbines [109]. This quantitative description of ice accretion provides valuable insights for icing studies and laboratory experiments. Aerones has developed a tethered UAV, a multi-rotor drone with long-duration hovering capability powered by tethered cables instead of batteries, which can spray deicing fluid through a pump system, offering a potential solution for deicing wind turbines. Gidinceanu further analyzed the economic and risk implications of such drone-based deicing implementations [110].

However, it is worth mentioning that a comprehensive understanding of the fundamental mechanisms of wind turbine icing requires more real-world icing data collected by UAVs and sensor systems across diverse environments and various wind turbine blade sizes. Current research on blade icing and de-icing using UAVs is still in its early stages, constrained by the limitations and developmental immaturity of UAV technology.

4.1.3. Strain and Vibration Analysis Based on Target Tracking

After analyzing and processing images or videos captured by unmanned aerial vehicles (UAVs), vibration characteristics and strain measurements can be derived, providing a dynamic quantitative analysis technology for wind turbine blade inspections. Given the low texture characteristics of wind turbine blades, placing standard markers on the blades is an effective method to improve detectability. Khadka et al. employed digital image correlation technology to detect blade displacement using manual markers and an octocopter UAV equipped with a stereo camera [111]. They further utilized dynamic splicing techniques to obtain detailed blade vibration characteristics [112]. They mitigated image blurring caused by UAV vibration through a combination of system design and hardware choices, including using a more stable octocopter UAV, a narrow aperture for greater depth of field, and data processing software that defines reference points to track the same point during blade rotation.

However, the use of small quadcopter UAVs often results in high noise levels, even when hovering. The unstable attitude of the UAV directly causes image blur and reduces image quality. Moreover, the placement of extra markers on the blades can be compromised by environmental contamination and wear, which degrades feature recognition and matching. This placement also increases inspection complexity and may affect the aerodynamic characteristics of the blades. In contrast, marker-free methods, while more flexible and natural in application, often require more advanced image processing techniques and adaptability to varying environmental conditions. To address these issues, Li et al. proposed using a high-pass filter and an adaptive scale factor to compensate for displacement drift [113]. They introduced a target-free discriminative scale space tracker (DSST) vision algorithm that accounts for background changes, enabling the recognition of blade natural frequencies from UAV videos captured with a single camera. Although practical applications still face challenges, the potential for non-contact acquisition of dynamic characteristics of wind turbine blades using UAV images is demonstrated in Figure 10 and Figure 11. These figures illustrate target tracking using digital image correlation (DIC) technology (marker-based) and the DSST algorithm (marker-free), respectively, but the techniques are not limited to these methods.

The direct use of UAVs equipped with sensors offers a promising approach to addressing the high failure rate of sensors deployed on wind turbines, eliminating the need for complex sensor system optimization. However, it is important to note that these methodologies are currently limited to laboratory environments for studying wind turbine models, with few practical applications in real-world wind farms.

4.1.4. Cracks, Debonding, Erosion, and Other Damage Detection Using Deep Learning

With the rapid advancement of artificial intelligence technology, particularly deep learning within machine learning, there has been a notable surge in research interest in computer vision and pattern recognition. Deep learning utilizes neural network structures akin to the human brain, particularly networks with multiple layers, to adeptly discern intricate patterns and features within data. Among deep learning methods, convolutional neural networks (CNNs) have received considerable attention for their exceptional performance in tasks, such as image recognition and classification. Through training these deep neural networks, models can autonomously learn feature representations from blade images, thereby enabling high-precision damage detection while enhancing efficiency and reducing labor costs. As computing power improves and big data technologies are increasingly applied, the performance of deep learning models in detecting damage on wind turbine blades continues to be refined and enhanced. This section examines the application of deep learning methods in the detection of blade damage, encompassing various aspects, such as dataset creation, data preprocessing, model architecture, training strategies, and performance evaluation.

In the wind energy sector, particularly in wind turbine blade damage detection, the most critical challenge is the scarcity of sample data. In many real-world industrial scenarios, defect samples are limited, often consisting of only a few dozen images. Because deep learning methods heavily depend on the quantity and quality of datasets, the lack of sufficient data significantly degrades model performance due to poor extrapolation capabilities for unseen damage. Consequently, deep learning for wind turbine blade defect detection primarily involves two approaches.

The first approach focuses on acquiring a sufficient number of images of wind turbine blades through field measurements, often utilizing data augmentation techniques, such as flipping [114] or random cropping [115]. There are also publicly available datasets that provide valuable resources for researchers.

Table 5. Comparison of wind turbine publicly available datasets.

Dataset	Purpose	Images	Environment	Source	Description
Blade-SfM [117]	Blade reconstruction	531	1	Outdoor laboratory shooting	Different distance 2/4/6 m; horizontal image overlap; vertical image angle change
Blade-Surface [118]	Blade reconstruction, damage detection, and roughness quantization	299	2	Indoor and outdoor laboratory shooting	Contain ground truth microscopy data Small blade segment with sand blasted
DTU [64]	Damage detection	701	1	UAV field measurement	The same wind turbine inspection images of two years
RDF [119]	Damage detection	431	1	UAV field measurement	Some of images being collected from DTU
Blade30 [51]	Blade reconstruction, defect detection, segmentation, classification, and deduplication	1302	6	UAV field measurement	Contain more regions, more weather, more environments, and more utility

details several such datasets, including Blade-SfM, Blade-Surface, DTU, RDF, and Blade30, which offer a range of images and annotations for training and evaluating deep learning models. Moreover, Xu et al. adopted a novel approach by generating virtual synthetic datasets at the height of wind turbines and simulating the environmental background [116]. This method provides an alternative data source that can complement field-measured images, particularly in scenarios where obtaining a large number of real-world images is challenging.

An alternative approach to deep learning for wind turbine blade defect detection involves pre-training models on general datasets. These datasets, typically used for training classification tasks, contain images of various common objects but do not specifically feature wind turbine blades. Utilizing transfer learning, the models are then fine-tuned with a relatively small number of wind turbine blade images.

Table 6. Deep learning models with transfer learning for damage detection of wind turbine blade surface based on UAV images (corrosion 1: rust, 2: pitting; Vortex Generator panel with Missing Teeth: VGMT).

Model	Pre-Training Dataset	Training Samples	Testing Samples	Detective Types	Accuracy (%)	Supplement
VGG16+PCA [121]	ImageNet	73,918	21,085	Damage or not	55.5F1	Manually remove background; unsupervised anomaly detection through One-Class Support Vector Machine; model compression through Principal Component Analysis
AlexNet-tl-rf [52]	ImageNet	900	450	Damage or not	98.49AA	Otsu threshold segmentation; random-forest-based ensemble learning classifier
Cascade Mask R-DSCNN [122]	COCO	650	51	Broken, crack, erosion, lighting damage, corrosion 1, VGMT	82.42MAP	Regular augmentations: flip, zoom, shearing; depth-wise separable convolution-based Resnet 50 and feature pyramid network
MobileNetv1-YOLOv4 [123]	PASCAL VOC	976	586 (390 validation samples)	Contamination, corrosion 2, crack, spalling	88.61MAP	15 kinds of augmentation operations; attention-based feature optimization using SENet: 92.94MAP, ECANet: 91.90MAP, and CBAM: 90.29MAP

Annotation: TP (the number of true positives), TN (the number of true negatives), FP (the number of false positives), FN (the number of false negatives). $A\left(\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}\right)={\displaystyle \frac{TP+TN}{TP+FP+TN+FN}};AA\left(\mathrm{A}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{e}\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}\right)={\int }_0^{1}a\left(r\right)dr;$ $P\left(\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\right)={\displaystyle \frac{TP}{TP+FP}};AP\left(\mathrm{A}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{e}\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\right)={\int }_0^{1}p\left(r\right)dr;$ $MAP\left(\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n}\mathrm{A}\mathrm{P}\right)={\displaystyle \frac{1}{N}}\sum A{P}_i;R\left(\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}\right)={\displaystyle \frac{TP}{TP+FN}};F1 Score=2\times {\displaystyle \frac{P\times R}{P+R}}$.

presents the performance of several models employing this transfer learning approach. The feasibility of using general deep learning models for surface damage detection on wind turbine blades and photovoltaic panels has been analyzed [120]. While supervised learning in deep learning models necessitates manual labeling, recent research has explored unsupervised learning models to automate the classification process. These unsupervised models require only category labeling at the end, significantly reducing human involvement.

The effects of utilizing various deep learning models, including CNN and its derivatives, for wind turbine blade damage identification are presented in Table 6. This table details the data volume, proportional relationship, damage types (with similar descriptions of the same damage unified across different articles), and detection effects achieved by each model. The models are organized into categories: the CNN series (including VGG and AlexNet), and the RCNN and YOLO series for target detection. Only the optimal results and their corresponding data volumes are presented for various dataset proportions, as different proportions may yield varying levels of accuracy. Accuracy figures, represented as percentages with varying decimal places, have been standardized from their original formats. The table’s footnotes clarify the differing evaluation indices used across models.

Due to potential variations in fault orientation and hardware performance, the specific complexity and operation time parameters of the optimized models are not included in this article. However, the methodologies employed for optimization are discussed in the following table. Furthermore,

Table 7. Deep learning models for damage detection of wind turbine blade surface based on UAV images (contamination 1: dirt, 2: oil dirt; debonding 1: coating defect, 2: gelcoat off, 3: navigational paint off; damaged lightning receptor: DLR).

Model	Training Samples	Validation Samples	Testing Samples	Detective Types	Accuracy (%)	Supplement
CNN [124]	1150	248	247	Damage or not,	94.94A	Augmentation through Keras neural network library containing flip, zoom, and shearing
	9505	3471	/	crack, erosion, lightning damage, mechanical damage, tip open	90.6A
Deep CNN [65]	5400	/	600	coating defect, crack, debonding 1, erosion, fiber defect	97A 97F1	Extended Haar-like feature extract; region proposals through Adaboost cascade classifier
VGG-11+ADMM [125]	20,422	/	5351	Contamination 2, debonding 2,3, erosion	92.8F1	Model compression through alternating direction method of multipliers
AlexNet [126]	10,000	/	6 × 350	Crack, sand holes	99.001AA	Test using 6-turn 350 images
Res-CNN3 [127]	1552	/	681	Crack, delamination, erosion, fatigue damage	80.6MAP	Res-Net; temporal channel complexity simplification; region proposals through selective search
Faster R-CNN [64]	726	/	433	Erosion, DLR, VGMT	81.10MAP	Augmentation through multi-scale pyramid, patching scheme, and regular; architecture: Inception-ResNet-v2
YOLOv3 [95]	600	/	146	Crack, erosion, DLR, VGMT	96MAP	Deblurring through super-resolution CNN
AFB-YOLOv5 [115]	2396	/	599	Damage, contamination 1	82.7F1 83.7MAP	Segmentation through random cropping; weighted bidirectional feature pyramid network; coordinate attention module; EIoU replace CIoU
SOD-YOLOv5 [114]	15,920	4549	2274	Contamination 2, corrosion, debonding 2,3	95.1MAP	Segmentation through Grab Cut and Hough transform; add a micro-scale detection layer; convolutional block attention module; computation reduction through channel pruning algorithm
MI-YOLOv5 [38]	819	102	102	Early crack	93.2MAP	Augmentation through slice transposition and reconstruction; architecture: Mobilenetv3, Ghostnet and Alpha-IOU; C3TR replace C3

showcases the effects of employing transfer learning with pre-trained models and their corresponding publicly available datasets. This table also outlines the damage types and detection effects achieved through this approach. The integration of pre-trained models with fine-tuned datasets specific to wind turbine blade images enhances the models’ performance and adaptability to this particular domain.

For different task requirements or regions of interest, it is necessary to establish specific evaluation criteria to measure model performance accurately. In classification tasks, accuracy is the most commonly used performance metric due to its simplicity and directness, reflecting the proportion of correctly classified samples to the total number of samples. In damage detection, the focus is often on the proportion of true damages to detected damages (P-precision) and the proportion of true damages to total damages (R-recall). To comprehensively measure the performance of precision and recall, the F1 Score, which balances the two metrics, is adopted. Averaging or integrating probabilities is a method to enhance the validity of the result without altering the fundamental performance measure.

The results presented herein demonstrate the feasibility of using consumer-grade UAV aerial photography images as data sources for wind turbine blade damage detection. This approach eliminates the need for higher-resolution image sources, making it a cost-effective and practical solution.

4.2. Thermography

While optical photography provides high visibility and is capable of detecting surface damage on wind turbine blades, thermal imaging using infrared cameras supplements this by offering additional temperature information. UAV inspections have demonstrated the ability to detect different types of damage from a single thermal video during wind turbine operation [128,129]. Figure 12 shows an example of damage recognition, classification, and regional location based on thermal images. It should be noted that not only these two types of damage can be identified but also flow separation, which is difficult to detect from optical images. Galleguillos et al. successfully detected delamination, impact, and cracks in blades of 10 m and 40 m using an industrial-grade intelligent unmanned helicopter [128]. Chen et al. further explored the practical application of this technology in different thermal backgrounds due to structural and camera movements, using a consumer-grade quadcopter [129]. They also established a publicly available dataset that includes both test and field measurements.

Despite the challenges of thermal imaging, which often exhibits weak texture and low contrast, these images can still be processed using deep learning architectures. Yu et al. proposed a U-Net segmentation model that utilizes geometric transformation parameters for extracting blade edges and stitching together entire blades [130]. For similar challenges with optical images, the preferable processing of thermal image stitching is shown in Figure 13.

4.3. Ultrasonic Inspection Robot and Embedded Acoustic System

It is to be noted that both optical and thermal cameras face significant challenges, as the quality of captured images is heavily influenced by weather conditions. Conversely, ultrasonic detection methods, though not affected by weather-related issues, suffer from efficiency constraints due to their limited detection range [25]. Given the inherent contradiction between the need for surface contact during ultrasonic inspection and the necessity of maintaining a safe flight distance, the use of unmanned aerial vehicles (UAVs) as direct inspection tools is not feasible. Instead, UAVs can serve as effective platforms for carrying negative pressure absorption robots.

Jiang et al. proposed a novel approach involving UAVs for precise navigation to predetermined positions [131]. At the designated location, the UAV determines suitable landing points with the aid of Light Detection and Ranging (LiDAR) technology. Subsequently, a negative pressure absorption robot is deployed and retrieved via a link hook mechanism, facilitated by a robot deployment interface and an on-load attaching module positioned on a horizontally laid blade. Sun et al. explored a different methodology, achieving relative static motion between the UAV and a service blade model [132]. This approach allowed for the deployment of a trolley onto the blade using a telescopic rod. The schematic structure and test of these two multi-robot systems are shown in Figure 14. These innovative techniques highlight the potential for UAVs to enhance ultrasonic inspection by serving as versatile platforms for deploying specialized inspection robots.

Acoustic emission, characterized by transient elastic stress waves generated by the release of energy from localized sources, offers a promising method for detecting issues in rotating components of wind turbines. García Márquez et al. presented an acoustic signal acquisition device mounted on a UAV equipped with an acoustic sensor [133]. They utilized wavelet transforms to analyze energy changes in engine rotation under the influence of noise. This method demonstrates the potential of acoustic emission techniques for effective monitoring and diagnostics in wind turbine maintenance.

5. Autonomous UAV Flight Control for Damage Detection

Flight operations involve the manual operation of UAVs, with pilots remotely controlling their movements. The blade damage detection studies described in Section 4.1 primarily involve manual flight operations, which are more suitable for preliminary small-scale or specific area inspections. The quality of the images largely depends on the pilot’s experience, and the plan—including the selection of viewpoints and the distance and path between the UAV and the wind turbine—must be adjusted according to the actual situation to meet operational requirements. However, to reduce human involvement and enhance autonomy, efforts have been made to develop autonomous UAVs. Autonomous inspections can continuously perform tasks irrespective of operator skill level and fatigue, improve efficiency and data consistency, and reduce labor costs for long-term operations.

The environment in a wind farm is generally static, with few new wind turbines or structures. For the flight planning of autonomous UAVs used in wind turbine blade damage detection, model-based methods are predominantly employed, while model-free methods like simultaneous localization and mapping (SLAM) are primarily applied in tracking and navigation within unknown environments. Model-based approaches necessitate prior knowledge, such as inputs from known Building Information Modeling (BIM) models or previous UAV flights. These autonomous systems are designed to meet the precise positioning and navigation requirements essential for wind turbine inspection. The flight control of multi-rotor UAVs involves several critical aspects, including wind turbine positioning, distance control, path planning, and coordination among multiple UAVs for collaborative operations. These functionalities are crucial for conducting efficient and effective wind turbine inspections.

5.1. Wind Turbine Localization

In manual flight, the location of the wind turbine is determined through visual observation and the display of the image transmission system, allowing for timely corrections of the UAV’s position. In contrast, for autonomous flight, the localization of the wind turbine is a prerequisite for inspection. Wind turbine localization involves accurately determining the position and characteristics of the detected turbines to establish the relationship between the captured images and the turbine model. This process ensures precise navigation and inspection during autonomous operations.

Typically, this process involves simplifying turbine components, such as the hub and blades, into points, circles, and line segments illustrated in Figure 15, using techniques like the Hough transform [97,134,135,136,137] and deep learning algorithms [138,139]. However, these methods often overlook the nonlinear curvature of the turbine blades. To address this limitation, Ma et al. employed Plücker and Cayley geometric methods to perform nonlinear optimization of line features [140]. This approach allows for a more accurate representation of blade curvature, enhancing the precision of wind turbine localization.

5.2. Distance Control between the UAV and the Wind Turbine

In the studies involving UAVs with fixed focal length cameras, precise control of the distance between the UAV and the wind turbine is essential for ensuring image sharpness. Maintaining an optimal distance is critical, as being too far or too close can significantly compromise data accuracy and safety. The error magnitude in suboptimal scenarios can be several times, or even tenfold, greater than in optimal conditions [141]. Meanwhile, the chosen distance directly impacts image processing efficiency, often eliminating the need for excessive detail in full blade regions. In manual flight operations, pilots typically select the preliminary distance based on an analysis of operational requirements and adjust according to the actual image quality.

To address this challenge, various sensor technologies, including Lidar and cameras, are employed to measure distances. Distance estimation is achieved using optical flow algorithms, random sample consensus, stereo triangulation, and other methods [140,141,142,143,144,145]. Subsequently, control strategies, such as proportional–integral–derivative controllers, often implemented through cascade control structures, are utilized to adjust the distance as required [143,144,145]. The control system is then integrated with navigation systems based on GPS and IMU for dynamic control and obstacle avoidance, ensuring safe and effective operation of the UAV during wind turbine inspection missions.

5.3. Path Planning of UAV Flight

After completing the aforementioned steps, the realization of autonomous flight for a single UAV necessitates the final and most critical step: path planning. Path planning involves determining the optimal route from a starting point to a destination within a given environment. This can be approached differently depending on various optimization objectives, such as shortest distance, minimum time [146], or minimum energy consumption considering wind effects [147]. A wind turbine typically has two or three blades of identical structure and shape. UAV inspection requirements entail not only the complete imaging of a single blade but also efficient coordination to the position of the next blade requiring inspection. Traditional S-shaped or circular paths, commonly used in mapping, are unsuitable for wind turbine blade damage detection. Therefore, specialized path planning is essential. Three types of UAV flight path planning are demonstrated in Figure 16.

Ivić et al. introduced a trajectory planning method based on a heat-equation-driven area coverage algorithm [148]. This method generates paths using potential fields and avoids obstacles by employing distance fields, assuming prior knowledge of the wind turbine’s three-dimensional structure. Zhang et al. reported that utilizing adaptive paths resulted in 2.7 times less error compared to pre-designed paths [141]. Real-time or adaptive planning relies on measurements from laser rangefinders and real-time analysis by the onboard computer. This approach avoids the need for pre-planned UAV navigation, enhancing robustness but increasing complexity and equipment requirements.

It should be noted that due to the limited operating range and duration of UAVs, it is sometimes necessary to transport the UAV to the designated location by boat or car to begin the inspection. Consequently, UAV path planning that focuses solely on the relationship between a single UAV and a single wind turbine may be accurate on a small scale, such as for an individual wind turbine, but is not necessarily optimal for an entire wind farm. Even when considering certain factors, such as wind influence and UAV endurance, the positional relationships between wind turbines within a wind farm are often overlooked.

Meanwhile, some researchers approach UAV inspection design as a traveling salesman problem (TSP) or a generalized TSP involving traversing viewpoints (selected checkpoints). For onshore wind farms, Baik and Valenzuela proposed a vehicle-mounted UAV model, as illustrated in Figure 17a [149]. This model first optimizes the UAV’s path within wind turbine clusters and then optimizes the vehicle’s path using integer linear programming, essentially treating the problem as two independent TSPs. In contrast, for offshore wind farms, Huang et al. proposed a boat-mounted UAV model, as shown in Figure 17b [150]. This model employs mixed integer nonlinear programming to minimize total energy consumption across four subproblems: the minimum detection time for a single wind turbine using a single UAV, the total detection time in the wind farm using a single UAV, the total flight time, and the total ship movement time. Both of two researches utilized the K-means clustering algorithm to determine that the entire wind farm is divided into several sub-regions containing different numbers of wind turbines. Additionally, the final clustering of wind farm based on geographical coordinates for truck-mounted UAV inspection path planning are demonstrated in Figure 17a (the assignment of serial numbers aims to wind turbines identification). However, Huang’s approach optimizes the UAV path based on boat clusters determined according to communication distance rather than the spatial arrangement of wind turbine clusters, alongside optimizing the paths of the load devices.

5.4. Coupling and Decoupling of UAVs and Robots

It is evident that single UAVs have limitations beyond visual inspection and simple non-destructive testing [151]. Mission efficiency, coverage area, and environmental adaptability are explicitly constrained by the UAV’s endurance and sensor performance. Multi-robot systems, incorporating multiple UAVs and robotic systems, directly alleviate these limitations. However, UAV-based multi-robot systems face challenges, such as task scheduling and coordination [152].

In multi-UAV planning, numerous factors must be considered, typically resulting in NP-hard problems, which means there is no known polynomial-time algorithm that can accurately solve all cases. Therefore, heuristics or approximate algorithms, such as genetic algorithms, simulated annealing, graph search algorithms, and reinforcement learning, are used to obtain feasible solutions for specific cases. Besides the single UAV planning problem in wind turbine damage detection, optimization factors include total flight time, battery exchange, and the number of UAVs required to cover a wind farm. Chung et al. proposed a heuristic algorithm for optimizing the placement of the minimum number of UAVs in wind farm inspections, taking into account the impact of wind on the range and speed of UAVs [146].

For the integration of UAVs and robots (typically one UAV paired with one robot), Jiang et al. designed an autonomous multi-robot system that addresses the coupling and decoupling of UAVs and detection robots [131]. This system employs an intelligent global mission planner and a mechatronics module to efficiently coordinate the tasks of UAVs and other robots. In this research, the UAV functions as a navigation and positioning tool to place and retrieve the robot at specified positions. Future research should consider interdisciplinary problems, such as task coordination, data transmission, and real-time decision making for multi-dimensional blade inspection robot systems.

6. Discussions

6.1. Major Challenges

With the advancement of unmanned aerial vehicle (UAV) technology, the application of UAVs in the wind power industry, particularly for the damage detection of wind turbine blades, has grown significantly. This technology offers advantages, such as high flexibility, mobility, improved economic benefits, and reduced safety risks. However, several challenges and limitations still need to be addressed to fully realize its potential.

• UAV performance: it is noteworthy that most UAVs reviewed in this paper are electrically powered, making them small in size with high maneuverability. However, this also introduces issues related to stability, reliability, and flight attitude requirements. The limited load capacity and space of these UAVs, often sacrificed for portability, restrict battery life and the weight and size of sensors. Furthermore, their operability in harsh environmental conditions, such as rain, snow, and strong winds, remains limited.
• Sensor performance: the performance of sensors mounted on UAVs is inherently limited by the platform’s constraints. Optical and thermal imaging cameras, although useful, cannot directly acquire three-dimensional data. While larger multi-rotor UAVs and unmanned helicopters can perform three-dimensional measurements using LiDAR, their application in wind turbine operations and maintenance remains relatively uncommon.
• Operators and governments: obtaining flight permissions for UAVs in both commercial and public wind farms poses significant challenges due to commercial sensitivity and security considerations. The lack of publicly shared wind farm data and the potential misalignment of data annotations with UAV inspection requirements further complicate this issue. Ensuring the privacy and security of wind farms necessitates strict regulations and procedures, which can hinder the efficient use of UAVs. On the other hand, the absence of standardized protocols and regulations for UAV inspections of wind turbines represents another challenge, potentially leading to inconsistencies in data collection, analysis, and interpretation, thus affecting the accuracy and reliability of inspection results. Developing standardized procedures and guidelines, along with training and certification for operators, are crucial for the widespread adoption and effective utilization of UAVs in wind turbine inspections.

The challenges in UAV-based wind turbine inspection primarily stem from detection equipment limitations and regulatory restrictions. Improving the endurance and load capacity of UAVs depends on advancements in battery technology, materials, and UAV design. Currently, carrying and replacing more batteries or selecting larger UAVs are the most feasible and effective ways to increase the operation time and load capacity of electric UAVs, though these solutions come with increased costs. Additionally, designing small UAV airports, centrally controlled by ground stations with pre-path planning, appears to be a viable scheme to enhance continuous operation capability and efficiency. For tasks beyond aerial photography (for which micro-UAVs with integrated cameras are relatively mature), larger UAVs are necessary to meet sensor load requirements. However, this entails a high initial investment cost for both the UAV and the sensors.

The rapid development of UAV technology has introduced various technical capabilities and application scenarios, but it has also raised a series of regulatory and ethical issues. Many countries have implemented relevant regulations, including UAV registration, flight height restrictions, and designated no-fly zones. Beyond establishing and perfecting a rigid regulatory framework through top-level design, sub-management units at all levels, and system improvements, it is essential to innovatively establish a wind industry standard for UAV inspection. This can be modeled after agricultural plant protection applications. A comprehensive approach would involve developing universal indicators based on different types of wind farms and operational task requirements, establishing complete processes, clarifying responsibilities and accountability mechanisms, and standardizing the format and content of inspection reports.

Meanwhile, moral and ethical concerns regarding the application of UAVs are equally important and currently lag behind regulatory developments. Given the requirement for sufficient wind resources, wind farms are typically located far from residential areas, thus minimizing privacy issues. However, for distributed wind turbines, including community wind turbines near residential areas, obtaining permission is necessary. Alternatively, the high accessibility of such wind turbines can be leveraged to adopt more suitable diagnostic methods. A fundamental solution to privacy concerns involves developing a robust legal system, which includes UAV registration, integration of manned and unmanned aerial vehicles, and clear identification of liability for infringements. This framework would clarify the subject of infringement and assign responsibility in cases of actual infringement. Regarding environmental impact, UAVs generally have a smaller footprint compared to wind turbines due to their smaller scale. This results in lower noise levels and reduced impact on aerial wildlife. While the resource consumption and influence of UAVs on the wind field are minimal, they still warrant consideration.

6.2. Prospects

The applications of artificial intelligence (AI) technology hold immense potential, particularly when supported by extensive datasets that facilitate optimal model optimization and accurate decision making. In the context of wind turbine inspection, unmanned aerial vehicles (UAVs) equipped with diverse detection equipment provide a unique opportunity to capture comprehensive multi-point time series data, primarily high-definition imagery, including wind vector measurements. Leveraging these capabilities to generate high-quality datasets and establish relevant industrial standards is essential. Moreover, openly sharing such datasets could significantly contribute to both scientific research and industry advancement, despite the inherent challenges.

Establishing three-dimensional models of wind turbines using structure-from-motion and multi-view stereo techniques has profound implications for UAV flight path planning, enhancing autonomy, spatially locating damage, and assessing conditions. Integrating AI techniques in a complementary manner can significantly bolster these capabilities. Generalizing AI models to various wind turbine environments enhances their applicability, while transfer learning reduces the need for extensive data acquisition, thereby enhancing model effectiveness. Concurrently, developing explainable AI techniques is crucial for ensuring transparency and credibility in damage detection and lifespan predictions, offering deeper insights into the mechanisms underlying damage and its impact on turbine longevity. Specific advancements include the correlation algorithm used to compensate for errors induced by displacement and rotation in UAV-based blade damage quantification, essential for achieving higher precision. Additionally, under conditions ensuring data quality and privacy protection, sharing datasets with annotations, usage instructions, and operational context is recommended. Optimizing the target detection model using foundational frameworks like YOLO or RCNN is indispensable, with exploration of unsupervised learning methods showing promise for further enhancement. Furthermore, developing 3D models of wind turbines based on UAV images facilitates precise path planning and three-dimensional damage visualization, alongside optimizing target assignment and global and local path planning for both single and multi-UAV autonomous operations.

In the meantime, the development of multi-robot systems, including multi-UAVs, UAVs with climbing robots, and UAVs integrated with ground vehicles or ships, holds significant promise for collaborative detection and real-time decision making. However, several technical difficulties must be addressed to achieve this. First, it is essential to solve the positioning and navigation challenges of multi-robot systems to enable effective collaboration. Second, robust communication between robots is required for real-time data exchange and decision synchronization. Third, optimizing path planning algorithms is crucial to ensure efficient coverage and minimize overlap. Finally, developing intelligent task assignment strategies is necessary to maximize the effectiveness of each robot in the system.

The conflict between the rapidly advancing wind energy sector and the limitations of traditional methods is becoming more pronounced. The evolution of UAV technology presents a promising solution. This review article comprehensively examines current research and applications of UAVs in detecting damage to wind turbine blades. It covers various aspects including types of typical blade damages, methods for detecting these damages with comparative analyses of their strengths and weaknesses, and the specific application of UAVs in this context. Furthermore, the article discusses challenges encountered, proposed solutions, and future prospects. It aims to offer valuable insights and perspectives to practitioners to advance structural health monitoring and enhance intelligence in the wind energy industry.