This paper aims to evaluate the implementation of Non-Destructive Inspection Techniques (NDIT) in the wind energy sector. For this purpose, a use case where AI-enhanced vision algorithms for anomaly detection in the painting inspection process in the wind energy sector is presented. Limitations and criteria for selecting the optimal hardware will be discussed, as well as the different parameters used for selecting, training, testing and validating machine vision applications in this field. Finally, the evaluation metrics of the algorithm used to evaluate the confidence level of the proposed model are explained, its performance on real, unseen data is presented, and future lines of action, as well as potential alternative applications, are summarized.
The use of Non-Destructive Inspection Technologies (NDIT), as opposed to traditional destructive
procedures, brings numerous benefits, including the competitive advantage of automating the manual
inspection processes currently used in many industrial processes [
In this context, the use of artificial intelligence (AI) systems allows to solve and automate
classification and prediction problems in the industrial environment. Integrating NDIT with Artificial
Intelligence (AI) applications is a zero-defect strategy to improve first-time right rates in production
environments [
New policies supported by the European Union based on the 2030 agenda and climate target plan
foster wind energy generation in the coming years [
Within this energy transition framework, current research focuses on meeting the demand for faster, more efficient, and more accurate quality inspection services in the wind energy industry that align with the zero defects and zero waste (ZDZW) framework. Implementing NDIT for real-time inspection can reduce material and energy consumption and lead times through higher inspection rates and a reliable automatic inspection process while improving overall operational costs. The increase in operational performance will be achieved through reduced energy and materials consumption, improved quality, and reduced labor thanks to automatic inspection solutions, making the EU's wind energy industry more resilient and competitive.
Integrating Artificial Intelligence (AI) into the European wind energy sector establishes a robust
connection with several other SDGs, underlining the multifaceted impact of this technological
advancement. The recent development of artificial intelligence (AI) in the European wind energy
sector aligns seamlessly with the Sustainable Development Goals [
By including non-destructive techniques and inspection algorithms within the European wind energy sector, process times and costs are reduced for the generation of resources that contribute directly to the European wind energy generation capacity (SDG 7 'Affordable and Clean Energy') and indirectly to the optimization of material and energy resources (SDG 13 'Climate Action') while encouraging the practice and integration of these systems that promote innovation in the industrial sector (SDG 9, Industry, Innovation and Infrastructure). It must be noted that integrating machine vision systems in industrial environments with harmful particles in suspension, as occurs in the painting process, increases safety in the inspection operations by mitigating the risks associated with manual labor while promoting the training and creation of new jobs related to AI (SDG 8 'Decent Work and Economic Growth'). Finally, collaboration between technology developers, energy producers, and policymakers is fostered thanks to AI integrations (SDG 17 ‘Partnership for the goals’). Partnerships forged through shared objectives, technological advancements, law-making and knowledge exchange are essential.
The deployment of automated inspection systems represents an innovation in the quality assurance policies in manufacturing steel towers for wind energy equipment, whose goal is the implementation of new approaches such as zero defect and zero waste methodologies. Currently, the paint inspection process is carried out manually by operators inside the painting cabin, so it ends up being a repetitive task based on the subjectivity of each operator and visual fatigue due to the inspection of large objects during significant periods. Implementing automated inspection systems reduces human fatigue associated with manual labor, ensuring consistent attention to detail throughout the painting process. The vision systems based on artificial vision improve reliability and accuracy in detecting anomalies, which are required to comply with Original Equipment Manufacturer (OEM) quality standards of the wind energy sector.
The efficacy of the AI-enhanced vision system has received considerable attention over the last
years [
The adequate neural network and model selection depends on the trade-off between
computational resources and accuracy requirements. It is within this framework where YOLOv8
capabilities are presented is a state-of-the-art object detection algorithm known for its real-time
processing capabilities and enhanced accuracy and speed in the detection of painting defects in
realtime applications [
Achieving higher levels of accuracy and reliability for Artificial Vision systems requires specific model training and high-quality data. The use case scenario dataset encompasses diverse scenarios, replicating real-world conditions along the manufacturing process. The model objective lies in detecting the six most recurrent defects within the painting process. Pinholes, blistering, inclusions, scratches, delamination and crumples are identified as critical defect classes, each with distinct characteristics and implications for the quality of the painted surface (Figure 1). The dataset utilized for the development and fine-tuning of the YOLOv8 model has undergone significant expansion throughout model iterations. Initially comprising 299 annotated anomalies, the dataset has grown to incorporate 401 annotated anomalies in its latest version, allowing the model to learn and generalize from a more diverse set of anomaly instances. However, more than 85% of the images used for training are anomalies of the inclusion type, resulting in a phenomenon known as sampling bias.
Once the data acquisition for the use case is performed, all these images are annotated using squared Bounding Boxes (BB) that enable the anomaly location within the image. To use the dataset in a suitable way, it is subdivided into three categories: train, test and validation using 70%, 20% and 10% of the images in the dataset, respectively (Table 2). The training hyperparameters selected for the current study determine a batch size of 16 images, 200 epochs, considering that the training stops if there is no improvement in the last 50 epochs. The remaining parameters have been set to their default value for the training process. At the hardware level, the resources that have been allocated for training correspond to NVIDIA RTX 3090 24 GB x1, 525.60.11 drivers & CUDA 12.0, MSI Z270 (MS7A63); 32 GB and Intel i7-7700K (4.2 GHz).
For the evaluation of the trained model in both iterations, it is represented by the Average Precision (AP) to measure the prediction accuracy and Intersection over Union (IoU) to measure the overlap between two BBs. An IoU limit over 50% is defined to determine whether a prediction is regarded as truth. For the presented use case, maximum mAP50 and mAP50-95 metrics achieved during training were collected and summarized (Table 3). The results show that an increase of 100 images with defects in the training set improved the mAP50 by 37%.
Once the training results of the customized YOLOv8nano model have been theoretically evaluated, its performance is assessed in an industrial production environment by inferencing images collected and never used in the training process. By evaluating the prediction results, the class anomaly, bounding box region, and confidence level can be observed, and all provide information on how specific the algorithm is to the provided prediction. For the proposed use case, inclusions are easily detected by the YOLOv8 model (Figure 2).
The results from the current AI algorithm are based on the neural network and trained with a dataset of 290 images, and the inclusion detection defect (Figure 1A) in the painting process is easily detected. The current trained algorithm allows the identification of inclusion anomalies, but it cannot correctly identify the remaining anomaly classes due to the low dataset employed. Further action will include the dataset augmentation with new high-quality data and improve the detection capabilities on the non-inclusion classes. 4. Conclusions
1. In this paper, it has been developed an artificial vision system to implement automatic inspection of painted surfaces of windmill towers. The model has proven good performance (0.37 mAP50) in detecting inclusions, the most common defect of the use case. Promising results in the aiming fully automating the task and reducing waste. Challenge of data collection and maintenance of production machine learning models. It is essential to consider that data collection for model training is a recurring task that must be carried out periodically to adapt our model to new types of defects and to balance the dataset so that all classes of anomalies are correctly represented. Increasing the number of annotated anomalies in the dataset contributes to the model's enhanced ability to detect and classify anomalies in diverse scenarios, including different lighting conditions, and ultimately improves the deployed solution's reliability and effectiveness. As future work, it is planned to extend the dataset samples to properly cover other types of defects.
The ZDZW project has received funding from the European Union’s Horizon Eu-rope program under grant agreement No 101057404. Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union. Neither the European Union nor the granting authority can be held responsible for them.
The author(s) have not employed any Generative AI tools.