Photovoltaic Cell Defect Detection Based on Weakly Supervised Learning With Module-Level Annotations

Hyungu Kang, Jeongmin Hong, Jiwon Lee, Seokho KangIEEE Access 2024 · OpenReview

Abstract

Recently, convolutional neural networks (CNNs) have proven successful in automating the detection of defective photovoltaic (PV) cells within PV modules. Existing studies have built a CNN based on fully supervised learning, which requires a training dataset consisting of PV cell images annotated according to whether the individual cells are defective. However, manually annotating all the PV cells is labor-intensive and time-consuming, leading to substantial annotation costs. In this study, we propose a weakly supervised learning method to build a CNN for cell-level defect detection in a cost-efficient manner. Our method uses a training dataset solely with module-level annotations indicating whether each PV module contains defective cells, thereby substantially reducing the required annotation costs. The CNN is trained in a weakly supervised manner such that all cells in a normal module are classified as normal and at least one cell in a defective module is classified as defective. The CNN can then be used to detect cell-level defects in new PV modules. The effectiveness of the proposed method is validated through experiments using real-world data provided by a PV module manufacturer.

Introduction

Quality control in solar manufacturing relies heavily on Electroluminescence (EL) imaging to identify defects such as micro-cracks, contamination, or "dead zones." While Deep Learning, specifically Convolutional Neural Networks (CNNs), has shown great promise in automating this detection, standard fully supervised models present a significant bottleneck: they require human experts to label every individual cell within a training set.

Given that a single PV module typically contains dozens or even hundreds of cells, the effort required to label thousands of modules at the cell level is immense. This research addresses the challenge of imprecise supervision: can a model be trained to identify a defective cell within a module using only a binary "pass/fail" label for the entire module?

Figure 1: Comparison of a defective PV module (containing a crack) with individual normal and cracked cells.

Methodology: Weakly Supervised Learning

The authors frame defect detection as a Multiple Instance Learning (MIL) problem. In this context, a PV module is treated as a "bag" and the individual cells are "instances."

Proposed Training Framework

Unlike conventional methods that require $N \times K$ annotations (where $N$ is the number of modules and $K$ is cells per module), the proposed method requires only $N$ annotations. By training the model to recognize that a "defective" module contains at least one faulty cell, the researchers achieved high-accuracy cell-level detection while drastically reducing manual labor.

Figure 4: The proposed weakly supervised learning framework, which trains a CNN using only module-level annotations.

Weakly Supervised Loss Function

The core of the approach is a specialized loss function designed with two distinct logic gates:

  • For Normal Modules: The model is penalized if any cell in the module receives a high defect score, pushing all cell scores toward zero.
  • For Defective Modules: The model is penalized based only on the maximum defect score among all cells in that module. This encourages the model to identify at least one cell as highly defective without requiring prior knowledge of which cell is faulty.

Experimental Setup and Results

The study utilized a real-world dataset from Shinsung E&G, a South Korean PV manufacturer, consisting of 1,418 module EL images (each with 144 cells).

Key Findings

  • High Diagnostic Accuracy: The VGG11 model achieved an AUROC of 0.9926, demonstrating near-perfect separation between normal and defective cells despite the lack of granular training labels.
  • Model Complexity: Interestingly, smaller models like VGG11 outperformed deeper architectures like ResNet101. The authors suggest that the sparse signal provided by weak supervision makes complex models more prone to instability or overfitting.
  • Localization Capabilities: Heatmap visualizations confirmed that the model correctly identifies specific defect types, such as cracks and dead zones, effectively bridging the "granularity gap" between module-level labels and cell-level diagnostics.

Conclusion

This weakly supervised approach allows manufacturers to leverage existing pass/fail data from assembly lines to train high-precision diagnostic tools. By eliminating the need for exhaustive manual labeling, this method significantly reduces the cost and time required to deploy automated quality control systems in the solar industry.

Summary updated

Abstract

Recently, convolutional neural networks (CNNs) have proven successful in automating the detection of defective photovoltaic (PV) cells within PV modules. Existing studies have built a CNN based on fully supervised learning, which requires a training dataset consisting of PV cell images annotated according to whether the individual cells are defective. However, manually annotating all the PV cells is labor-intensive and time-consuming, leading to substantial annotation costs. In this study, we propose a weakly supervised learning method to build a CNN for cell-level defect detection in a cost-efficient manner. Our method uses a training dataset solely with module-level annotations indicating whether each PV module contains defective cells, thereby substantially reducing the required annotation costs. The CNN is trained in a weakly supervised manner such that all cells in a normal module are classified as normal and at least one cell in a defective module is classified as defective. The CNN can then be used to detect cell-level defects in new PV modules. The effectiveness of the proposed method is validated through experiments using real-world data provided by a PV module manufacturer.
BibTeX
@inproceedings{kang2024photovoltaic,
  author = {Hyungu Kang and Jeongmin Hong and Jiwon Lee and Seokho Kang},
  title = {{Photovoltaic Cell Defect Detection Based on Weakly Supervised Learning With Module-Level Annotations}},
  booktitle = {IEEE Access},
  year = {2024},
  url = {https://openreview.net/forum?id=6KCcasnkvl},
}
Figure