Deep Learning-Based Automatic Micro-crack Inspection in Space-Grade Solar Cells

Abstract

Spacecraft power systems reliability is critical parameter for mission success. Multiple checks and inspections are carried out for each component for power subsystems. In this paper, a novel automated solar cell micro-crack inspection tool is presented which is based on convolutional neural network (CNNs) to classify space-grade multi-junction solar cells taken under electroluminescence condition. The whole system is named ELSIS, which stands for “Electroluminescence Smart Inspection System”. It is an end-to-end automated system that acquires images under electroluminescence condition as arrays, identifies each cell and classifies them into two classes namely the cells that exhibit micro-cracks and cells and those that are free of micro-crack. ELSIS is developed with the objective to reduce the tedious and time-consuming manual effort required to identify micro-cracks and to increase the reliability of the solar cell modules by minimizing the human errors which may arise during the manual inspection of thousands of solar cells for spacecraft by automating the whole process. The CNNs for ELSIS have been developed in TensorFlow framework in Python based on InceptionV3 architecture. ELSIS is augmented with the latest image processing techniques that are applied to acquire EL images, which are modules containing array of solar cells (Fig. 1). ELSIS thus identifies and produces individual solar cells from the arrays and indexed and stored. The deep learning network was trained on a large number of solar cell >6000 images such that cross-entropy of the network settles within an accepted constant value. The trained network when tested for a large sample size of ~3000 new solar cells yielded a very reliable >98% accuracy. ELSIS helps in capacity building for scale manufacturing of satellites with stringent power budget, making it imperative to have zero defect solar cells.