Defect Detection in Solar Cells: Leveraging Deep-Learning Technology

Abstract

The purchase of a solar panel can be considered as a long-term investment, especially for solar projects. The industry norm is to provide decades long warranties on a solar panel's power output, which means that the performance dropping below the promised minimum during this period would equal a substantial cost to the manufacturer. To combat this, the manufacturers employ a handful of different testing and inspection methods. One of the methods that has been gaining popularity in recent years is electroluminescence testing, and is often combined with machine learning. Most of the research with electroluminescent images focuses on making models that can act as an automated defect detector, without taking into account the differences in prediction errors.

In this thesis, we explored the possibility of making defect detection models that takes the severity of the prediction errors into account, while also providing an effective method for handling uncertain predictions with manual inspections. A custom weighted loss-function was made in an attempt to force the model to make less "risky" predictions, and was compared up against performance with a standard loss function. We utilized two datasets, one with images labeled based on probability of being defective, and one which labeled images based on defects present. In an attempt to incorporate a model which could learn from both datasets and predict both types of labeling, we employed a specialized training strategy over several steps. Class activation mappings were also compared up against bounding boxes which were provided for one of the datasets, to find if this could be a useful tool for detecting the defects.

We found that the models were substantially better at predicting the defect types rather than if it was defective or not. The training process also showed that the datasets were not compatible with each other, making the combined model unsuccessful. Using the custom weighted loss-function, we managed to impact the predictions of the model, making it worse overall while it made less severe errors. The class activation mappings were able to mostly highlight the relevant defect areas, with the exception of the thick-line and dislocation defects.

The results indicate that it is possible to force the model to consider the cost of the different prediction errors, though the model and the loss-function would require more work to be usable in practice. The class activation mappings would be able to provide assistance in manual-inspection, but a more specialized defect localization model could be a better alternative. While our thesis made frequent use of estimates and approximations, any further work should include proper estimates and data from the solar industry to check the validity of the findings.