Electroluminescence technology is enabling the inspection of PV panel components that are not visible to the naked eye, or detectable in other forms of solar panel inspections, write Dr Michelle McCann and Lawrence McIntosh from PV Lab Australia, with Liam O’Duibhir, Carsten Eckelmann, Aaron Tranter and Yan Wang
Electroluminescence testing is one of the techniques at the heart of our solar panel testing business at PV Lab Australia. This testing allows us to see deep inside a solar panel to view things we can’t see with the naked eye and that can’t be seen with other tests.
Electroluminescence (EL) is used by system owners and insurers to look for damage after an extreme weather event; by system owners or builders to look after their investment by checking regularly for unexpected degradation of panels at large solar farms; by builders or installers to check for damage during panel installation; and by purchasers to check for damage during shipment.
Electroluminescence is done in two parts. The first task is to capture an image of a solar panel. The second part is to analyse this image.
Image capture is essentially done by operating the solar panel in reverse: applying a current and getting light out, similar to an LED since a solar cell is also like a diode. It’s pretty easy to do this in the dark, either at night for a large solar farm or in a dark room in a laboratory, but methods also exist for daytime electroluminescence.
In the old days…
For a large solar farm, the most rudimentary method of image capture is to drive around at night with a camera mounted on a tripod in the back of a ute with a power source. The power source can be used to electrify whole strings or individual panels (as required) and the camera captures an image. Some minutes are required for each panel.
Image analysis is done after image capture and is desk-based work that takes some minutes for each panel. A person is required to zoom into the image, focusing on just one cell at a time to assess whether it has any features of note. If it does, these are noted in a spreadsheet. This process is then repeated for the 60 to several hundred cells in the panel.
Drone-based inspections
In order to speed up image capture at PV Lab Australia, we have partnered with Quantified Energy Labs (QE Labs) to bring drone-based electroluminescence imaging to Australia for the first time.
QE Labs is based in Singapore and is a spinoff from the National University of Singapore. The company’s autonomous, drone mapping technology is up to 20 times faster and 10 times cheaper than tripod-based electroluminescence. In a recent project that also happens to be one of the world’s largest EL inspections, they imaged 120,000 modules in a floating PV plant in just two weeks.
The system provided by QE Labs is very hands-off. Using the site map as input, the flight path for the drone is autonomously planned and the drone flies on autopilot. Images captured from the drone can be georeferenced to allow reliable identification of any defective modules.
However, speed rarely comes for free and there is a slight drop in image quality compared to tripod-based inspections. Depending on the panel fault being detected, this may not be relevant, but in cases where it is relevant, the ability to scan an entire solar farm and then narrow in on panels or regions of interest more than adequately compensates in terms of time and money spent.
Photo at top of page: QE Labs’ autonomous drone on an electroluminescence inspection run. This drone solution is up to 20 times faster and 10 times cheaper than manual electroluminescence inspection. Photo: PV Lab Australia.
Machine learning
It’s no use capturing images of 120,000 modules across two weeks if you need minutes to analyse each one. Even at just one minute of analysis per panel, 120,000 minutes is almost three months of analysis time.
Enter machine learning (ML). Machine learning is particularly well suited to the task of image analysis as solar panels contain distinct and homogeneous cells, allowing automatic, fast and accurate detection of defects.
Image recognition using machine learning has demonstrable success in many other industries such as medicine, advanced manufacturing and satellite imagery. Additionally, these algorithms can continually improve and monitor their accuracy, drastically reducing the need for human intervention.
Machine learning algorithms can also scale to demand, increasing the throughput of the current process exponentially. Accurate and automated detection of defects would significantly decrease analysis time and costs to the customer.
In order for a machine-learning approach to be successful, three things are vital:
- A large enough data set to accurately train the system.
- Smart algorithms.
- Models that are built and trained to run at real time in what is effectively a production setting.
To this end, PV Lab Australia has teamed up with two local companies, Aqacia and 2pi Software. This work is supported by Innovation Connect in Canberra.
Aqacia provides machine learning and artificial intelligence (AI) solutions for high-tech industries, taking the most difficult real-world problems and crafting solutions to enable seamless integration of AI into production and research settings.
2pi Software is a specialist AI/ML enablement company providing a digital platform that reduces the heavy lifting and software complexity involved in high-volume data ingress and egress to and from neural networks. This frees data scientists to focus on the core domain problem being addressed.
For the project, PV Lab Australia brings a data set of many tens of thousands of panels and quadsquillions of thousands of cells that have been analysed individually. Aqacia is developing automatic image processing and ML inference models to automate defect identification, and 2pi Software will build a solution that takes advantage of the high-performance computing and parallelisation capabilities of the Amazon Web Services (AWS) cloud to allow flexible data processing workflows to be constructed to optimally train AI/ML models.
This partnership is ultimately supportive of the drone work. Depending on the size of the data set, analysis by humans may take weeks, whereas the AI technology can process hundreds of panels in minutes.
The faster speed means panels can be assessed in real time in both the PV Lab production environment and in an onboard drone environment. Real-time assessment enables fast feedback to system owners and iterative control of the inspection process if a more detailed assessment is required in certain situations.
“The 2pi Software team is greatly honoured to be involved in such a groundbreaking PV sector innovation and the application of emerging AI/ML techniques to aid monitoring of solar panels’ ongoing health and lifetime efficiencies – a truly foundational aspect of the promise of PV,” says 2pi software director Liam O’Duibhir.
Bringing it all together
This work is almost (but not quite) too good to be true. The final product achieves the trifecta of being:
- Faster, thanks to drones and machine learning.
- More accurate because human error is reduced thanks to machine learning.
- Cheaper for the customer – thanks again to drones and machine learning.
This future is nothing to be fearful about. Sure, it involves machine-learning enhanced drones with night vision on a search-and-destroy mission to weed out underperformance at solar farms, but we think it is a good thing.