Defect Detection in Wind Turbine Blades Using Infrared Thermography, Image Processing, and U-Net ^†

Abstract

In this research, we developed and tested an automated defect detection system for wind turbine blades using infrared thermography (IRT) and the deep learning model U-Net.

1. Introduction

The inspection of wind turbine blades [1,2] is very important for ensuring both the productivity and safety of wind energy systems. In this paper, we present the development of an automated defect detection system that uses infrared thermography (IRT) to extract the database and the U-Net deep learning model to identify defects in wind turbine blades. We began with a theoretical analysis using COMSOL 6.3 simulations to study the best time of day for capturing infrared images in order to maximize the visibility of defects based on thermal conditions. Our second analysis allowed us to settle the limits of defect sizes in relation to their depths that could be detected using infrared thermography. The simulated defects were air cavities of different sizes, which were analyzed to evaluate the ability of our system in detecting those defects at different depths. Next, we used the U-Net deep learning model for the automatic detection of defects in the wind turbine blade infrared images. We used U-Net for its pixel segmentation of the defects, which provided accurate identification of the defect areas. Additionally, we employed image processing techniques in order to divide the images corresponding to different blades of the turbine. Then, we compared the images of the blades at the same positions using masks to improve the reliability of the defect detection. The findings of this work show that the proposed system successfully identifies defects of different sizes in turbine blades. However, this detection system faces a challenge with identifying false positives mistaken for real defects caused by some external factors like blade contamination and variable sunlight reflections due to clouds. Despite these limitations, the system provides a potential solution for automatically identifying defects in wind turbine blades, which improves productivity and decreases the need to perform manual visual drone inspections.

2. Thermal Modeling of the Wind Turbine Blades

For the simulation process, a wind turbine blade model was used, including a full-scale blade with dimensions of 60 m in length, 5 m in height, and 2 m in width, as shown in Figure 1. Our blade model contained four subsurface defects of different sizes.

3. Results of the Simulations

In order to study the best timing for capturing infrared images, we executed different simulations for the month of September. These simulations provided a sequence of infrared images that clearly showed the appearance of (simulated) defects, as shown in Figure 2. After this, we extracted the temperatures of the defect areas and the surface temperature from 10 consecutive points throughout the day, from 14:30 to 20:00, with measurements taken every 30 min. The choice of this time range was based on the solar radiation power being the most ideal for defect detection during these hours. The results of these simulations are illustrated in Figure 3. The simulations were performed in order to replicate a sunny day, and for the temperatures of the defect areas, we calculated the average temperature of the four defects simulated.

4. Field Image Acquisition

Following our preliminary study and the availability of equipment, we conducted fieldwork at a wind farm south of Quebec City on Thursday, 12 September 2024. The weather was cloudy in the morning and sunny in the afternoon, though with some clouds, which enabled infrared image capture between 16:00 and 18:00, and this was timed to align with the optimal conditions identified in our COMSOL study. We used a cooled FLIR (Wilsonville, OR, USA) model A6701 MWIR camera with a 100 mm lens, as shown in Figure 4.

We achieved a clear view of several sections of the wind turbine blades. During this test, infrared image sequences were recorded for two wind turbines.

5. Methodology–Image Processing

We developed several image processing techniques in order to isolate and locate possible defects on the infrared images of the two wind turbine blades we inspected. By employing these methods, we were able to clean the infrared images of the clouds in the background in order to prevent the detection of these clouds as defects and to improve the visibility of the defects. This methodology served to guarantee the correct study of different regions of the blades. We started by converting the raw infrared images to a binary format to feature the defect regions, followed by some morphological operations such as erosion and dilation in order to refine the images by eliminating noise and restoring the defect regions. Next, we applied some image processing methods for identifying and grouping similar images to split the sequences of the images. The images were divided into sub-lists corresponding to specific blade sections (Blade 1, Blade 2, and Blade 3) for localized defect analysis. Furthermore, we operated a comparison of images from various blades (Blade 1, Blade 2, and Blade 3) at the same location using multiple thresholds to assist us in identifying defects on the blades. The creation of the combined defect masks helped us in detecting minor variations between the images. This resulted in more uniform defect detection in the dataset [3].

6. Automatic Defect Detection Using U-Net

The automatic defect detection method using U-Net was successful in detecting many defects that had already been detected by the image comparison technique where the blades were compared at the same positions, as shown in Figure 5. Detection validation was performed with another approach [3]. The defects detected by both methods were considered as true defects. In contrast, the defects detected only by one method and, more specifically, those not validated by U-Net were considered false positives. This dual validation approach helped us in filtering out the errors caused by environmental conditions, like sunlight reflections or debris, reducing the chances of misinterpreting non-defect areas as actual defects. This was a preliminary analysis, and the network could be refined and better exploited.

7. Conclusions

In this paper, we developed and tested an automated defect detection system for wind turbine blades using infrared thermography and the deep learning model U-Net. We began our study with a theoretical study using COMSOL simulations in order to determine the best time of the day for capturing infrared images and the limitations of the defect detections in terms of size and depth.

Our approach included two main detection methods: image comparison across turbine blade images [3] and U-Net segmentation. U-Net allowed us to achieve precise defect localization by segmenting the infrared images pixel by pixel. This approach, mixing image comparison and deep learning, allowed us to cross-validate the detected defects, confirming (possible) true defects and identifying false positives.