Automated flaw detection in images is becoming increasingly important in industrial quality assurance. Whether in manufacturing, logistics or infrastructure, visual inspection is indispensable in many sectors. However, as production volumes rise, manual visual inspection is reaching its limits. As a result, significant defects remain undetected or are only identified on a random basis.
But what if machines could learn to see what experts see – and do so with consistently high quality and on a large scale?
This is precisely what we have achieved in a joint project with CCE Holding, an international operator of renewable energy plants. The principle is transferable: AI-supported flaw detection in images can be used wherever visual data is crucial for quality and safety.
Visual inspection plays a key role in the manufacture of photovoltaic (PV) panels. Electroluminescence (EL) images – a type of X-ray image – can be used to detect micro-cracks, cell breaks or contact flaws that would not be visible to the naked eye.
However, with shipments of tens of thousands of panels, manually evaluating all EL images is simply not feasible. CCE Holding was therefore looking for a solution that would enable automated image analysis using artificial intelligence – with at least the same level of precision as manual inspection.
Together, we developed an AI approach that automates the entire visual inspection process – based on real industrial image data. An overview of the process:
1. Image segmentation into individual cells
Using traditional computer vision techniques, the panel images were divided into individual cells. This allows even the smallest structural components to be analysed in a targeted manner – a crucial advantage for the trainability of the AI model.
2. Data annotation by experts
A specially developed tool enabled subject matter experts to mark faulty cells quickly and intuitively. This resulted in a high-quality, realistic training dataset – combining human expertise with an efficient workflow.
3. Training and model architecture
Based on the latest research findings, we selected a suitable model architecture for visual classification. This resulted in an 87 % detection rate with just 2 % false positives – an impressive figure for an AI solution in industrial quality control.
The project is impressive not only for its precision but also for its applicability under real-world production conditions. It demonstrates that AI reduces the workload on humans and systematically improves quality assurance.
Outside the solar industry, too, there are numerous areas where visual quality inspection plays a role. Whether it be printed circuit boards in electronics manufacturing, surface inspections in mechanical engineering or packaging lines in the food industry. Wherever image data is already available, AI can be used to provide support.
The key lies in the combination of:
This case study demonstrates that automated flaw detection using AI is more than just a research project – it is an industrial reality. When technology is implemented in a way that is easy to understand, trained using specialist knowledge and applied in a practical manner, it creates measurable added value.
Do you have any questions or need assistance?
Anyone facing similar challenges should consider: Where can AI support visual inspection processes? Where can we save resources without compromising on quality?
We’d be happy to help answer these questions.
Would you like to find out more about artificial intelligence or our customers’ success stories? You’ll find an overview of all our Champion Stories on our website.