Barely-Visible Surface Crack Detection for Wind Turbine Sustainability

TL;DR

The paper tackles the challenge of detecting barely-visible hairline cracks on wind turbine blades, which are difficult to identify with conventional inspections and existing ML approaches. It introduces the Zeitview Crack Detection Dataset (ZVCD), a geographically diverse, real-world collection of high-severity crack imagery, and presents an end-to-end pipeline that tiles high-resolution images, performs classification on patches, and uses GradCAM-based region proposals to guide human review and maintenance decisions. A comparative study of lightweight CNN backbones (e.g., ResNet-18, EfficientNet-B3, MobileNetV3 variants) trained with a turbine-level 90/10 split demonstrates convergence and emphasizes precision to minimize missed cracks, with emphasis on on-board deployment feasibility. The resulting workflow supports scalable, automated wind-turbine inspections with a human-in-the-loop for validation, offering practical impact by reducing downtime and the risk of catastrophic blade failure. Future work includes exploring infrared-RGB comparisons and moving toward fully autonomous, scalable inspection pipelines.

Abstract

The production of wind energy is a crucial part of sustainable development and reducing the reliance on fossil fuels. Maintaining the integrity of wind turbines to produce this energy is a costly and time-consuming task requiring repeated inspection and maintenance. While autonomous drones have proven to make this process more efficient, the algorithms for detecting anomalies to prevent catastrophic damage to turbine blades have fallen behind due to some dangerous defects, such as hairline cracks, being barely-visible. Existing datasets and literature are lacking and tend towards detecting obvious and visible defects in addition to not being geographically diverse. In this paper we introduce a novel and diverse dataset of barely-visible hairline cracks collected from numerous wind turbine inspections. To prove the efficacy of our dataset, we detail our end-to-end deployed turbine crack detection pipeline from the image acquisition stage to the use of predictions in providing automated maintenance recommendations to extend the life and efficiency of wind turbines.

Figures (4)

• Figure 1: Geographic locations of the inspections in our ZVCD dataset. Our dataset consists of a diverse set of imagery of barely-visible surface cracks acquired from 4,684 turbine inspections from nearly 1,000 unique sites across the world.
• Figure 2: Validation accuracy performance during training of different models. We observe that all the models have converged well and show similar performance.
• Figure 3: Sample Crack Detection Region Proposals using the GradCAM method selvaraju2017grad. Our pipeline provides region proposals to a human analyst who then reviews and further diagnoses barely-visible surface hairline cracks for severity categorization and repair recommendations. We first compute the GradCAM attributions w.r.t the crack detection class model output. We then post-process the attribution heatmap using the normalize and polygonize operations.
• Figure 4: Our turbine crack detection pipeline diagram. Our pipeline consists of automated acquisition of overlapping imagery of each turbine blade using DJI UAVs and our flight planning algorithms. Each image is then decomposed into 1024 x 1024 tileswhich we then run our classifier on. Next, we compute and postprocess the gradient attribution heatmap on the high confidence predicted tiles. Our analysts are then shown these proposal regions containing cracks for review and in-depth analysis for inspection report generation.