Download the poster
windeurope.org/summit2018#GlobalWind2018
When it comes to using an Unmanned AerialVehicle (UAV) to collect imagery for windturbine inspections, maneuvering a drone
around a huge turbineisn’t the only challengeinvolved. There's alsothe issue of capturingextra positionalinformation so imagescan be projected ontoa plane, which isnecessary since thecamera is not
positioned to look directly down as in mostUAV imagery applications. Then there'sensuring the quality of the capture while inthe field to avoid the costs associated withhaving to re-fly a job. Finally, there's the timeand rigor required to analyze all the data toidentify and locate damage and otherabnormalities on turbine blades.
Wind turbine blades affected by leading edge erosion with (left) leading edge
delamination, (right) pits and gouges.
Sareen1. , A., Sapre, C. A. & Selig M. S. (2014): Effects of leading edge erosion on wind turbine blade performance. ̶ Wind Energy 2014; 17:1531–1542. DOI: 10.1002/we
Lusk, R. M. & Monday, W. H. (2. 2017): An early survey of best practices for the use of small unmanned aerial systems by the Electric Utility Industry. ̶ Oak Ridge National Laboratory, ORNL/TM-2017/93.
Ehrlich, M., Justice, B. & Baldridge B. (3. 2016): Harris Deep Learning technology. ̶http://www.harrisgeospatial.com/portals/0/pdfs/12-16_deep_learning_whitepaper_FA.pdf [Accessed 12 July 2018]
We present a workflow currently in use thattakes UAV imagery from flights around windturbine blades, and automatically assessesthose images for defects.
This method for wind turbine inspectionsignificantly reduces the cost of wind turbinemanagement and inspection, and allows toinspect more turbines in less time.
With an accuracy of 95+ %, there is littlevariation between these automaticinspections as compared to inspections donemanually by human technicians.
Future work on the project includes addingmore damage types, and assigning damageseverity classifications based on standards setforth by the Electric Power Research Institute(EPRI).
These enhancements will further informturbine managers on how to prioritize therepair and maintenance of their windturbines.
Using UAV Imagery and Deep Learningfor Wind Turbine Inspection
1Barrett M. Sather, 2Joshua M. Riedy & 3Thomas Bahr1Harris Geospatial Solutions Inc., 2EdgeData, 3Harris Geospatial Solutions GmbH
PO.250
Introduction Methods – Capture Assurance Results
Objectives
Conclusions
References
A web application, the BladeEdgeSM Portal,provides user friendly access to wind turbineblade inspection information generated bythe aforementioned workflow. Specifically,imagery is organized by turbine, blade, andblade side with anomalies and damagegraphically represented.
From this application users have thecapability to modify findings as well asorganize and output inspection results.
The two main metrics that we use to validateif a capture is good or not, and what they helpwith preventing are:
• Image overlap: ensures uniformity of flight,eliminates gaps, ensures mosaics arecontinuous.
• Pixel spacing: ensures uniformity of flight,flags cases where the data doesn’t match,distance to the blade (too far/close).
On-site validation of image metadata, GPS information, image overlap andpixel spacing using the BladeEdgeSM Capture Assurance Tool (BECAT).
www.harrisgeospatial.com
Multiple images are captured along the windturbine for all four sides, and all three bladesof a wind turbine. The current procedure is tofly one side of the blade at a time. Four flightsare required for a single blade. Blades areflown from root to tip, and the root and tiplocations are captured from directlyunderneath the feature.
Methods – Image Capture
Methods – Damage Detection
Damage detection is performed on the pixeldata using Harris Deep Learning technologies,and damage locations are reported for eachimage. Current models are trained on chipand crack damage types.
Labeling of imagery for the generation of models of blade edges, variousdefects, and background using Harris Deep Learning technologies.
Blade edge, background model, and resulting heatmap of leading edgeerosion, showing location and type of the defects.
The Deep Learning technologies are achieving95-99 % accuracy when identifying damageon wind turbine blades.