Wind turbines often operate in remote and harsh environments, making it difficult to detect, analyze, and manage defects such as cracks, corrosion, and overheating. This study addresses that challenge by proposing a novel approach that combines multi-sensor unmanned aerial vehicle (UAV) inspections with deep learning-based defect recognition and a fuzzy logic approach to assess criticality. The main problem lies in translating raw UAV-collected data into reliable and quantifiable metrics for defect severity, a gap that has inhibited timely interventions and thorough maintenance strategies. This study aims to improve the speed and accuracy of detecting and ranking anomalies in wind turbine blades, towers, and motor assemblies. Experiments show that our proposed multisensor pipeline, which merges thermal and visual data via an ensemble of convolutional neural networks (CNNs), raises detection accuracy by up to 13.5% compared to single-sensor or single-model approaches. Furthermore, the resulting fuzzy-based numerical scores align closely, within 0.15 deviation, with expert judgments of defect urgency. Our conclusions highlight the value of cross-channel data fusion and robust logic-driven rankings for wind energy applications. This approach contributes to reduced operational downtime, improved turbine longevity, and enhanced safety by minimizing overlooked damage and emphasizing prompt and effective maintenance. In summary, the study illustrates how automated UAV data acquisition, ensemble CNN detection, and a systematic fuzzy logic scoring mechanism can work to fill the existing gap in defect criticality assessment on wind power plants.
Wind power plants form a key segment of modern renewable energy. Their efficiency and reliability,
however, directly depend on the structural integrity of components such as blades, towers, and motor
assemblies. These parts endure continuous mechanical loads, varying weather conditions, and
corrosion-inducing environments [
The goal of this study is to enhance UAV-based inspections by proposing a comprehensive approach to determining the criticality assessment of defects on components of wind power plants detected by UAV sensors. The objective includes developing and validating an approach that fuses sensor data, ensemble deep learning, and fuzzy logic to produce interpretable severity scores for each defect. To this end, we identify three major contributions: • • •
An end-to-end approach that processes multispectral UAV data and detects multiple defect types via an ensemble of CNN architectures.
A structured fuzzy logic subsystem that computes a final numeric criticality score ( final) using both physical and thermal parameters, plus expert weighting.
A validated demonstration of the approach’s effectiveness through experiments comparing standard vs. composited data, and single-model vs. ensemble learning, yielding numerical insights into accuracy gains and improved recall rates.
The remainder of this manuscript is organized as follows. Section 2 surveys related works focusing on deep learning–based wind turbine inspections and fuzzy-logic criticality evaluations. Section 3 details the proposed approach to determining the criticality assessment of defects on components of wind power plants detected by UAV sensors. Section 4 provides a rich overview of experiments, including quantitative comparisons, tables, and figures showcasing how multispectral composition boosts detection metrics. Section 5 outlines advantages, disadvantages, and future research questions related to sensor calibration and real-time implementation. Finally, Section 6 concludes with a forward-looking summary of the major findings, numerical improvements, and pathways for extending this work.
Growing interest in UAV-based defect detection for wind power plants has led to numerous research
contributions. Traditional methods employed single-sensor red, green, and blue (RGB) data combined
with classic machine learning classifiers [
With the advent of deep learning, convolutional neural networks (CNNs) proved more effective
in capturing intricate, data-driven features from imagery. According to [
To further boost detection precision, ensemble methods combine multiple CNNs specialized in
different defect dimensions or spectral data [
Beyond finding defects, attention has shifted to evaluating criticality. Early threshold-based methods, e.g., labeling cracks above 1 cm as “dangerous,” proved oversimplified [20]. Fuzzy logic • • • • • • • might be a solution for bridging numerical data (e.g., defect size, thermal anomalies) with domain knowledge [21]. For instance, in [22], fuzzy membership functions quantify geometry, temperature, or curvature parameters, enabling nuanced severity rankings. While fuzziness better mirrors realworld uncertainty, few fully integrated UAV pipelines incorporate a fuzzy approach to ranking detected defects [23]. This gap underscores the relevance of the proposed approach to determining the criticality assessment of defects on components of wind power plants detected by UAV sensors,” as it merges ensemble CNN detection with a fuzzy criticality subsystem.
Based on the survey, this study addresses the overarching issue of translating multi-sensor UAV detection outputs into a structured criticality assessment for each identified defect. To achieve this, several tasks must be completed:
Task 1: Acquire and preprocess UAV data from multiple spectral channels, ensuring alignment between RGB and thermal images.
Task 2: Implement robust ensemble CNNs to detect diverse defect types under real-world conditions.
Task 3: Extract physical metrics (length, area) and thermal indicators (min/max/avg temperature) to characterize each defect.
Task 4: Incorporate fuzzy logic or a similar interpretive mechanism to generate numeric criticality scores reflecting domain expert knowledge.
These tasks guide the development of the introduced approach, described in detail below.
In this section, we detail the proposed approach to determining the criticality assessment of defects on components of wind power plants detected by UAV sensors. The approach integrates three primary stages: (i) data collection and composition from multiple UAV sensors, (ii) ensemble CNNbased defect detection, and (iii) fuzzy logic–driven criticality assessment.
Block 1: Physical Defect Characterization—Captures raw bounding boxes from the ensemble output, corrects distortions, and computes physical dimensions plus temperature metrics. Block 2: Expert Functions for Critical Defects—Formalizes cracks, corrosion, and overheating severity via specialized formulas, referencing the tables that assign weighting coefficients (β, γ, η) for each WTC.
Block 3: Fuzzy Integration and Final Scoring—Transforms each detected defect’s parameters into fuzzy sets, merges them with expert severity functions, and applies defuzzification to obtain final( ).
Block 1 implements an iterative process to transform detected bounding boxes into real-world dimensions and temperature profiles.
Step 1.1: Region of Interest (ROI) Extraction. For each detected defect, we crop its region from the undistorted image as follows: ROI = crop undistorted, ensemble, ensemble + e nsemble, ensemble + ensemble , where ensemple, ensemple is the bounding box’s top-left corner, and ensemple, ensemple are its width and height. physical units in the following way:
Step 1.2: Image Enhancement (Bilateral Filtering, CLAHE). Noise reduction and local contrast enhancement follow [24], preserving edges while boosting defect visibility.
Step 1.3: Adaptive Thresholding and Morphology. Binary segmentation via adaptive thresholding
separates the primary defect region (employed in our previous work [
Cracks: • •
Corrosion:
Overheating: where γ represents the coefficient of corrosion in towers and typically has a higher weighting factor than in mechanical joints.
= ⋅
, then follow.
where is distance to the defect, p is pixel size, and f is the camera focal length. Physical dimensions r eal, real ,
,
Step 1.5: Thermal Parameter Extraction. We also record minimum min, maximum max, and
average avg temperatures across the defect’s mask. Combined, these data form the complete model
complete used in subsequent blocks, which first was introduced in our previous work [
Based on the WTC (blade, tower, and motor), we apply distinct mathematical models to incorporate
expert knowledge regarding defect impact. We define three main functions for cracks exp( rift),
corrosion exp( cor), and overheating exp heating [
visible( , , )| ′( )| 1 + ( ) , exp( cor) = γ ⋅
( , , ) , Ω( ) Ωheating( ) exp heating = ⋅
Δ def( , , ) ⋅ |∇2 ( , , )| , where η reflects how critical temperature deviations are for motors, generators, or control electronics.
yield a single numeric criticality final( ).
The final block merges objective measurements (Block 1) with expert-based functions (Block 2) to
Step 3.1: Fuzzy Membership Functions. Each parameter ∈ complete (e.g., r eal, avg) is t-norm or minimum [26]. assigned a trapezoidal membership ( ) [25] based on expert-defined intervals.
Step 3.2: Parameter Aggregation. The fuzzy set ( ) combines these individual memberships via between the data-driven set and expert set.
Step 3.3: Expert Score as Fuzzy Set. The expert assessment exp( ) is transformed into a fuzzy set exp( ) using a Gaussian membership function [25] centered on exp( ).
Step 3.4: Consistency Coefficient. A cosine similarity measure ( ) quantifies agreement Step 3.5: Weighted Combination. Weights and are derived via a sigmoid function, balancing data-based membership with the expert membership.
Step 3.6: Fuzzy Aggregation is calculated as follows:
final( ) = D ⋅ D ( ) + exp ⋅ exp( ).
Step 3.7: Defuzzification is performed as follows: final( ) = ∫ ⋅ final( )
∫ final( ) .
(6) (7) specific severity thresholds.
The final crisp value final( ) thus integrates measured geometry, thermal data, and domainThis study was performed at a wind energy test site with steep terrain and frequent temperature swings. Turbines from a leading manufacturer, rated at 2–3 MW and equipped with rotor diameters over 100 m, were examined (illustrated in Figure 2). conditions made it suitable for testing adaptive UAV techniques.
Strong winds reaching 14 m/s challenged both UAV flight stability and our data acquisition strategy, ensuring realistic operational conditions. The experimental phases are presented as follows: • • • •
Phase I: Single-Sensor Trials. The system was tested using only the RGB camera from UAVA, establishing baseline.
Phase II: Thermal Addition. We introduced IR data from UAV-A to create composite images, measuring changes in detection accuracy.
Phase III: UAV-B Cross-Validation. We compared IR readings from UAV-A and UAV-B in overlapping flight patterns to confirm sensor calibration consistency.
Phase IV: Full Ensemble + Fuzzy. The final integrated pipeline (RGB + IR + ensemble CNN + fuzzy logic) was evaluated on newly acquired flight data, with real-time or near-real-time inference tested in a partial field environment.
We deployed two UAV platforms. The first was a DJI Matrice 300 RTK UAVs [27] carrying Zenmuse H20 (RGB) and H20T (thermal) cameras, whereas the second was a custom hexacopter fitted with a FLIR Duo Pro R. Both incorporated RTK receivers for centimeter-level positioning. Overlapping flight paths, planned via mission control software, captured images from altitudes of 10–70 m and standoff distances of 3–8 m relative to blades or tower surfaces. Manual ground inspections, including chalk-marked crack lines, served as references.
In total, 500 RGB and IR images were gathered per cycle, and each cycle was repeated thrice for statistical reliability. Raw data covered various vantage points and speeds, mitigating motion blur and ensuring thorough coverage. All image processing and CNN training were carried out on a dedicated server with dual RTX 3090 GPUs. This rigorous setup allowed us to validate the stability and performance of our method under actual wind farm conditions, revealing its feasibility for ongoing turbine inspections.
We used multiple indicators to assess both detection performance and alignment with expert-labeled severity. First, Accuracy measured the overall proportion of correct predictions, while Precision and Recall evaluated the quality of positive classifications and the rate of correctly retrieved defects, respectively. F1-score combined Precision and Recall into a single measure for balancing false negatives and false positives. Further details on these metrics are comprehensively discussed in the recent survey by Rainio et al. [28].
Area under the Curve (AUC) captured threshold-independent detection capabilities by comparing true positive and false positive rates across varying cutoffs. For criticality alignment, we correlated the fuzzy-based final with domain expert severity scores, reporting Pearson’s r and Spearman’s ρ to determine consistency.
We also examined runtime efficiency on a GPU-based server, comparing inference speeds between YOLOv8 alone and its ensemble variants. This analysis accounted for bounding-box aggregation and the minimal overhead of fuzzy membership calculations to confirm applicability for real-world UAV deployments.
This section demonstrates, via comprehensive experiments and numerical analysis, how the presented approach advances the state of the art in detecting and evaluating defects on WTCs.
Here, we describe our experiments on various UAV-collected images of blades, towers, and motor compartments. Each category exhibits distinct challenges, as summarized in Table 1. Notably, motor assemblies require careful thermal analysis for detecting incipient overheating, while blade surfaces demand sub-millimeter cracks resolution.
corrosion cracks benefit from thermal changing weather data Cracks, High, integrates well with corrosion color vs. thermal contrast Cracks, High, offsets perspective corrosion distortion with IR data Corrosion Medium, simpler geometry
but high altitude Overheating, Critical, highlights localized corrosion hotspots in IR
Vent channels, strong reflections
Figure 3 demonstrates examples of defect detection results on both (a) standard RGB and (b) composite imagery using YOLOv8 and ensemble methods.
(a) (b)
Notably, composite images facilitate identifying subtle cracks in shaded regions and pinpointing thermal anomalies that might escape pure RGB inspection.
We compare YOLOv8 [15] alone and in ensemble with RetinaNet [29], EfficientDet [30], and Cascade R-CNN (CR) [17]. Table 2 shows how combining YOLOv8 with Cascade R-CNN yields the highest overall recall (93.0%–94.0%), especially on composited multispectral images.
Defect type
Cracks
Model
YOLOv8 + RN YOLOv8 + ED YOLOv8 + CR
YOLOv8 YOLOv8 + RN YOLOv8 + ED YOLOv8 + CR
YOLOv8 YOLOv8 + RN YOLOv8 + ED YOLOv8 + CR
Table 3 aggregates average metrics across all defect classes to evaluate the overall detector performance. The results in Table 3 confirm that YOLOv8 + Cascade R-CNN stand out, with about a 4.7% improvement in accuracy over YOLOv8 alone.
Composited RGB + IR images significantly augment detection outcomes vs. standard RGB alone. Table 4 highlights how Accuracy and AUC improved across cracks, corrosion, and overheating classes, with an especially dramatic jump (over 20% in some cases) for overheating detection.
Accuracy rose from 78.8% to 92.3%, while F1-score climbed from 75.7% to 89.5%. This underscores the benefit of incorporating thermal signatures, especially for subtle or hidden damage.
We further evaluated the fuzzy logic subsystem’s alignment with expert judgments. Table 6 shows 10 example defects, with measured physical/thermal parameters, the expert criticality rating, and the fuzzy computed final.
The difference in final scores (0.15–0.2) from expert-labeled values validates the approach’s strong consistency. In-depth calculations, such as the cracks example with length = 1.2 m and curvature = 0.05, show how the system’s fuzzy logic approach yields final ≈ 4.8, closely matching the expert’s 5.0 rating. Similarly, corrosion or overheating examples maintain small error margins (< 0.02), confirming reliability.
Compared with earlier research on UAV‐assisted WTC inspections, our approach achieves higher
detection accuracy and improved interpretability. Previous single‐sensor or machine learning
methods (e.g., YOLO alone or SVM‐based classifiers) often reached an 80% accuracy limit under
varying illumination, whereas our ensemble solution, which merges thermal and RGB data, surpasses
90% across all defect types. This outcome agrees with prior multi-sensor findings indicating that data
fusion and advanced deep learning enhance detection performance [
A key advantage of our pipeline is the interpretability that emerges from the fuzzy logic subsystem. While some earlier works adopt threshold-based severity triggers, those fixed thresholds cannot adapt to dynamic or context-specific variations in turbine components. The fuzzy approach, on the other hand, draws upon membership functions to incorporate both geometric and thermal data, creating continuous numeric outputs that refine the classification of “low,” “moderate,” or “high” risk anomalies. This adaptiveness allows the method to capture subtle changes that static thresholds would ignore.
However, there are some disadvantages. One major limitation is increased computational overhead. Ensemble detection, particularly in scenarios where YOLOv8, Cascade R-CNN, and other CNNs each produce bounding boxes for merging, can require more powerful hardware than singlemodel solutions, which may be impractical for lightweight UAVs that only have minimal onboard processing. Another disadvantage is the fuzzy subsystem’s reliance on expert weighting. When domain knowledge is limited or contradictory among experts, membership functions may be miscalibrated. This can lead to inaccurate or unstable severity scores, especially for newer or evolving defect types not covered in the initial model.
In general, these findings point toward the method’s strong potential in everyday wind energy asset management yet also indicate challenges that remain open to further investigation. Among the limitations, calibration stands out: sensor offsets must be consistently monitored to avoid drift in geometric or thermal readings. Weather constraints also impose a natural boundary on flight times and data capture quality. A final set of research questions revolves around real-time edge computing, where UAVs would run the pipeline onboard, sending only summarized results to ground stations. Overcoming limited GPU resources on smaller UAV platforms is therefore both a technological and an algorithmic challenge.
This study introduced and validated an approach that combines UAV-based multisensor data acquisition, ensemble CNNs, and a structured fuzzy logic system to assess the criticality of defects on wind power plant components. Experimental trials confirmed that the fusion of thermal and visible-spectrum data, processed through YOLOv8 in tandem with other state-of-the-art detectors like Cascade R-CNN, increased detection accuracy by an average of 13.5% relative to single-sensor or single-model baselines. F1-scores rose by an additional 4.7%, indicating more balanced performance across multiple defect types such as cracks, corrosion, and overheating. Critically, the fuzzy logic subsystem assigned each detected anomaly a severity score final that diverged from expert-labeled judgments by just 0.15 on average, thus demonstrating reliable alignment with domain knowledge. Despite these improvements, the study also highlighted key challenges in sensor calibration, weather-dependent flight stability, and higher computational requirements imposed by ensemble detection.
Future research can extend the architecture with more advanced data streams such as LiDAR or 3D point clouds, refine fuzzy memberships for nuanced environmental conditions, and explore synergy with predictive maintenance, potentially enabling real-time UAV analytics for large-scale offshore or mountainous wind farms. Ultimately, this approach represents a significant step toward automated, accurate, and transparent WPP defect management.
The authors are grateful to the Ministry of Education and Science of Ukraine for its support in conducting this study, which is funded by the European Union's external assistance instrument for the implementation of Ukraine's commitments under the European Union's Framework Program for Research and Innovation “Horizon 2020.”
The authors have not employed any Generative AI tools. [13] A. Omara, A. Nasser, A. Alsayed, M. R. A. Nabawy, Remote wind turbine inspections: Exploring the potential of multimodal drones, Drones 9.1 (2024) 4. doi:10.3390/drones9010004. [14] J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: Unified, real-time object detection, in: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, New York, NY, USA, 2016, pp. 779–788. doi:10.1109/cvpr.2016.91. [15] G. Jocher, A. Chaurasia, J. Qiu, Ultralytics YOLOv8, 2023.
URL: https://github.com/ultralytics/ultralytics. [16] S. Ren, K. He, R. Girshick, J. Sun, Faster R-CNN: Towards real-time object detection with region proposal networks, IEEE Trans. Pattern Anal. Mach. Intell. 39.6 (2017) 1137–1149. doi:10.1109/TPAMI.2016.2577031. [17] Z. Cai, N. Vasconcelos, Cascade R-CNN: Delving into high quality object detection, in: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, IEEE, New York, NY, USA, 2018, pp. 6154–6162. doi:10.1109/CVPR.2018.00644. [18] Y. Mao, S. Wang, D. Yu, J. Zhao, Automatic image detection of multi-type surface defects on wind turbine blades based on cascade deep learning network, Intell. Data Anal. 25.2 (2021) 463– 482. doi:10.3233/ida-205143. [19] O. Melnychenko, O. Savenko, P. Radiuk, Apple detection with occlusions using modified YOLOv5-v1, in: 2023 IEEE 12th International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS), IEEE, New York, NY, USA, 2023, pp. 107–112. doi:10.1109/IDAACS58523.2023.10348779. [20] M. Patil, T. Abukhalil, S. Patel, T. Sobh, UB swarm: Hardware implementation of heterogeneous swarm robot with fault detection and power management, Int. J. Comput. 15.3 (2016) 162–176. doi:10.47839/ijc.15.3.849. [21] H. Zhu, J. Liu, H. Zhu, D. Lu, Z. Wang, A novel wind turbine fault detection method based on fuzzy logic system using neural network construction method, IFAC-PapersOnLine 53.5 (2020) 664–668. doi:10.1016/j.ifacol.2021.04.157. [22] L. Dubchak, A. Sachenko, Y. Bodyanskiy, C. Wolff, N. Vasylkiv, R. Brukhanskyi, V. Kochan, Adaptive neuro-fuzzy system for detection of wind turbine blade defects, Energies 17.24 (2024) 6456. doi:10.3390/en17246456. [23] M. H. M. Ghazali, W. Rahiman, A novel fault detection approach in UAV with adaptation of fuzzy logic and sensor fusion, IEEE/ASME Trans. Mechatron. 30.1 (2024) 381–391. doi:10.1109/tmech. 2024.3393990. [24] S. M. Pizer, E. P. Amburn, J. D. Austin, R. Cromartie, A. Geselowitz, T. Greer, B. ter Haar Romeny, J. B. Zimmerman, K. Zuiderveld, Adaptive histogram equalization and its variations, Comput. Vis., Graph., Image Process. 39.3 (1987) 355–368. doi:10.1016/s0734-189x(87)80186-x. [25] L. A. Zadeh, Fuzzy sets, Inf. Control 8.3 (1965) 338–353. doi:10.1016/s0019- 9958(65)90241-x. [26] S. Alshattnawi, L. Afifi, A. M. Shatnawi, M. M. Barhoush, Utilizing genetic algorithm and artificial bee colony algorithm to extend the WSN lifetime, Int. J. Comput. 21.1 (2022) 25–31. doi:10.47839/ijc.21.1.2514. [27] Support for Matrice 300 RTK - DJI, 2020. URL: https://www.dji.com/support/product/matrice-300. [28] O. Rainio, J. Teuho, R. Klén, Evaluation metrics and statistical tests for machine learning, Sci.
Rep. 14.1 (2024) 6086. doi:10.1038/s41598-024-56706-x. [29] T.-Y. Lin, P. Goyal, R. Girshick, K. He, P. Dollar, Focal loss for dense object detection, IEEE Trans.
Pattern Anal. Mach. Intell. 42.2 (2020) 318–327. doi:10.1109/tpami.2018.2858826. [30] M. Tan, R. Pang, Q. V. Le, EfficientDet: Scalable and efficient object detection, in: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, New York, NY, USA, 2020, pp. 10778–10787. doi:10.1109/cvpr42600.2020.01079.