Renewables inspection imagery analysis with computer vision

Using machine computer vision algorithms to identify defects in renewables assets

Research Review 2018

GTR has developed an automated image processing algorithm to identify and classify turbine damage and wear from drone captured inspection imagery. Utilizing the latest advances in convolutional neural networks and GPU cloud processing, DNV GL has proved the efficacy of the direct application of machine learning based artificial intelligence to renewables inspection processes.

We can only see a short distance ahead, but we can see plenty there that needs to be done.

Alan Turing

Computing machinery and intelligence

DNV GL forecasts a 65-fold increase of solar PV capacity (to 19 TW) and 15-fold increase of wind capacity (to 7.2 TW).

With such an increase of renewable assets, comes a need for improved monitoring to ensure continued reliability and safety of the electrical energy distribution system.  Owners and operators are increasingly adopting the use of drones to inspect assets gathering visual, performance and condition information. This large amount of data combined with the expertise of DNV GL renewables makes for the unique opportunity to apply computer vison machine learning algorithms to the task of reviewing and annotating inspection information. 

Traditional wind turbine inspections by rope access field personnel takes proximately 3 hours per turbine and generates a few kB of data. The collected images – limited to suspected damage and technician notes - take weeks of labour-intensive desk processing time by skilled engineers to identify and flag damage. On the other hand, a typical drone flight lasts 15 minutes per wind turbine inspection, recording visual video footage, optional hyperspectral or thermal imagery of the entire asset, as well as metrological (meteorogical?) conditional and flight information generating many GB of data to be analysed. Existing analysis tools are insufficient to accommodate these volumes of additional information. In order to capture additional insights the rich data offers, DNV GL has advanced the automation of damage detection using machine learning techniques from recent advances in computer vision. By utilizing pattern recognition capabilities of convolutional neural networks, a tool has been developed taking seconds to run. The AI tool produce predictions suggesting areas damage for an engineer to review. The entire asset integrity can be quickly documented with additional traceability of damage history logged.

Artificial Intelligence can be a supplement to human insight, not substitute.

Abhijit Naskar

GTR Power and Renewables have developed a recurrent U-shaped convolutional neural network (rUNet) trained on in-house wind turbine inspection imagery which detects damage in visual images. By stacking multiple CNN’s, this supervised machine learning technique utilizes the wealth of data from high resolution imagery, the latest in computer vision research and advances cloud base GPU processing on Veracity. The approach initially developed internally on electroluminescence imagery of solar voltaic cells from the DNV GL PVEL lab has now been refined and proven on wind turbine blade imagery. This achievement, taking algorithms developed on images taken from a controlled laboratory environment and adapting it to wind turbine field inspections data taken from varied environmental conditions, is a major  step in making machine learning research a usable tool for the renewables industry. Continued refinement incorporating thermal and hyperspectral imagery means extension of applications to inspections of installed solar production projects and enhanced conditions monitoring. As outlined in the 2017 DNV GL position paper Making Renewables Smarter, a variety of models and machine learning techniques will provide the support for artificial intelligence tools in wind and solar. This AI computer vision research provides DNV GL and our customers with the tools to excel in the digital age.