State-of-the-Art Deep Learning Anomaly Detection Method for Analyzing Electroluminescence Images of Solar Cells

Regarding the rapid growth of the solar energy industry, it is essential to update and automate its quality control and defect detection processes. Various methods and technologies are used to inspect solar modules and solar cells for detecting defective ones. One of the techniques to detect defects such as micro cracks or finger failures in solar cells is a visual inspection of their electroluminescence (EL) images. Considering an expert's required cost and knowledge to inspect EL images, an automatic mechanism for this process is more efficient. In this paper, we applied PatchCore, a state-of-the-art deep learning (DL) method of anomaly detection in images, to identify defective cells from their EL images, and its performance in detecting anomalies has been analyzed. In comparison to other DL methods, PatchCore requires a smaller training dataset, therefore it is more flexible and efficient for use in environments such as the production of new cell types. First, a training dataset is given to a deep neural network and in the test phase each image is given to the same deep neural network and its extracted features are compared to the normal features which are saved in a memory bank. Any deviations further from the normal state will be detected as a high abnormality score. Three different EL image datasets from laboratory and production site cells were created and different experiments were conducted to evaluate the PatchCore performance. Experimental results showed PatchCore achieved an AUROC score of 0.97 for detecting anomalies in a dataset of polycrystalline EL images. The overall experiments showed that PatchCore can be used as a high-performance anomaly detection method in EL image inspection processes.

Automated summary

In response to the rapid growth of the solar energy industry, updating and automating quality control and defect detection processes is essential. This study applied PatchCore, a state-of-the-art deep learning method, to detect anomalies in electroluminescence images of solar cells. Experimental results showed that PatchCore achieved an AUROC score of 0.97 for detecting anomalies in a dataset of polycrystalline EL images, indicating its high performance as an anomaly detection method in EL image inspection processes.