Anomaly Detection and Automatic Labeling for Solar Cell Quality Inspection Based on Generative Adversarial Network

Abstract

Quality inspection applications in industry are required to move towards a zero-defect manufacturing scenario, with non-destructive inspection and traceability of 100% of produced parts. Developing robust fault detection and classification models from the start-up of the lines is challenging due to the difficulty in getting enough representative samples of the faulty patterns and the need to manually label them. This work presents a methodology to develop a robust inspection system, targeting these peculiarities, in the context of solar cell manufacturing. The methodology is divided into two phases: In the first phase, an anomaly detection model based on a Generative Adversarial Network (GAN) is employed. This model enables the detection and localization of anomalous patterns within the solar cells from the beginning, using only non-defective samples for training and without any manual labeling involved. In a second stage, as defective samples arise, the detected anomalies will be used as automatically generated annotations for the supervised training of a Fully Convolutional Network that is capable of detecting multiple types of faults. The experimental results using 1873 Electroluminescence (EL) images of monocrystalline cells show that (a) the anomaly detection scheme can be used to start detecting features with very little available data, (b) the anomaly detection may serve as automatic labeling in order to train a supervised model, and (c) segmentation and classification results of supervised models trained with automatic labels are comparable to the ones obtained from the models trained with manual labels.

Few-shot incremental learning in the context of solar cell quality inspection

“…We have explored changing the architecture from the original proposal which has resulted in more precise segmentation. This technique adds another step in the methodology presented in [2] which completes the process to develop an inspection system by first using defect-free samples to generate an anomaly inspection model, then which will serve as automatic labeling to train a more accurate model, and now update the final model by incorporating new defect classes with few samples.…”

“…A possible alternative to avoid the need for defective samples may be to approach the problem from an anomaly detection perspective [1,2,3], where instead of creating the training set from defective samples, the set is generated using defect-free samples that are more accessible in a production line. However, as the objective of the unsupervised approach is to detect anomalous patterns which encompasses defective and non-defective patterns, it usually reports lower detection rates than supervised models that are strictly trained to detect only defective features.…”

“…In this work, we experiment with the incremental few-shot learning proposed in [15] in the industrial context of solar cell inspection where we try to update a model that has been trained on three base defect classes (i.e., cracks, microcracks, and finger interruptions) and incorporate new defects (i.e., black spots and bad soldering) such that the final model is able to segment five different defects. In this way, we will add another feature into the proposed anomaly detection based methodology [2] incorporating an adaptation stage for the supervised model to make it able to detect new types of defect classes that can arise during production, or also, incorporate features that were not previously considered important but for now on it should be controlled.…”

Detection of microcracks and dark spots in monocrystalline PERC cells using photoluminescene imaging and YOLO-based CNN with spatial pyramid pooling

“…In this case the autoencoder not only allowed the anomaly in PV modules be segmented, but it also facilitated defect detection. In another work a CNN based on generative adversarial network has been evaluated for detecting anomalies in solar cells [18]. Although tested on different datasets, such a network performs less satisfactorily with recall, precision and specificity averaging at 79%, 73% and 73% respectively.…”

A Novel Data Augmentation Method for Improved Visual Crack Detection Using Generative Adversarial Networks

“…However, this work is based on the fact that the GAN model is not trained on defected images, so it will reconstruct distorted images when the input contains defects, providing less control of the generation. In [20] the authors deploy a GAN-based model for detecting defects and anomalies on solar cell manufacturing. Although a novel approach for such a task, this work doesn't exploit the capabilities of GAN models for generating new samples, which is where GANs excel at.…”