ncpre-text Logo 2

Let the ‘Machine’ do your Electroluminescence image analysis

Does anyone accept the solar module with cracked cells? No! The cracks on the solar cells may cause a gradual reduction in power output. These cracks are usually invisible to the naked eye.

Here Electroluminescence (EL) imaging comes to the rescue, where one can see the cracks present in the cells. Unfortunately, the problem doesn't end here. An expert eye is required to detect and report these cracks in the cells. In addition to that, as the solar modules are produced and installed in bulk, the manual observation of such a large number of EL images is a tedious and time-consuming process. Due to the wide variety of visual features in EL images, reporting the presence of cracks may include human bias as well. So why not automate this process!

background image
Figure: Process flow for binary classification of cells

Advanced Image Processing and Artificial Intelligence (AI) techniques are becoming popular as a solution to such problems. Manual observation of cracks in EL images requires a person to be trained with a variety of EL images having different kinds of cracked and non-cracked features on the cells. Similarly, here in NCPRE, we have developed a deep learning model that is trained with large amounts of EL image data consisting of different features of cracked and non-cracked cells for automated classification of cells. The EL images were collected from installed PV plants from different climatic zones of India during the 'All India Survey 2018'. Unlike most of the published works in this area that have used the lab simulated data with artificially generated cracks, our data consists of the images of the fielded PV modules having cracks mainly due to environmental causes, poor installation practices, and transportation. Also, the camera used for capturing EL images was developed by NCPRE at a relatively low cost and is robust for field applications.

For cell wise analysis of EL images, it is important to crop out the background of the PV module and separate each individual cell. Various image processing algorithms were implemented in a sequence to prepare our cell EL images from the module EL images. The cracked and non-cracked cells were manually labeled, just like a textbook is prepared for our model to read and get 'trained' for the 'test'! ResNet50 is the deep network architecture used for the training. The parameters of the model are tuned in such a way that it will learn better to classify the cells with higher accuracy. The trained model is now prepared for the test on the image data which is not part of the training data set. The accuracy of the model in the test was found to be 98.6%! Similar experiments are performed with EL images from another source to study the performance of models on a cross dataset and it is found that active learning of models over new labeled data is the way towards a generic classification model.

Prof. Narendra
Prof. Anil
Prof. Rajbabu