Automatic Classification of Defective Photovoltaic Module Cells in Electroluminescence Images (1807.02894v3)

Abstract: Electroluminescence (EL) imaging is a useful modality for the inspection of photovoltaic (PV) modules. EL images provide high spatial resolution, which makes it possible to detect even finest defects on the surface of PV modules. However, the analysis of EL images is typically a manual process that is expensive, time-consuming, and requires expert knowledge of many different types of defects. In this work, we investigate two approaches for automatic detection of such defects in a single image of a PV cell. The approaches differ in their hardware requirements, which are dictated by their respective application scenarios. The more hardware-efficient approach is based on hand-crafted features that are classified in a Support Vector Machine (SVM). To obtain a strong performance, we investigate and compare various processing variants. The more hardware-demanding approach uses an end-to-end deep Convolutional Neural Network (CNN) that runs on a Graphics Processing Unit (GPU). Both approaches are trained on 1,968 cells extracted from high resolution EL intensity images of mono- and polycrystalline PV modules. The CNN is more accurate, and reaches an average accuracy of 88.42%. The SVM achieves a slightly lower average accuracy of 82.44%, but can run on arbitrary hardware. Both automated approaches make continuous, highly accurate monitoring of PV cells feasible.

• The paper introduces automated defect classification in EL images of PV cells using SVM with hand-crafted features and CNN with transfer learning.
• The CNN approach achieves 88.42% accuracy, outperforming the SVM method, especially on polycrystalline modules.
• These methods enhance solar module inspections by reducing reliance on expert manual analysis and lowering maintenance costs.

Overview of Automatic Classification of Defective Photovoltaic Module Cells in Electroluminescence Images

The paper investigates the implementation of automated methods for detecting defects within photovoltaic (PV) modules using electroluminescence (EL) imaging. While EL imaging offers high-spatial-resolution insights into potential defects in PV modules, traditional manual inspection remains inefficient and dependent on expert knowledge. This research seeks to automate the detection of these defects, enhancing efficiency and consistency in the maintenance of solar energy systems.

Methodology

Two distinct approaches are evaluated for the task of defect classification in EL images:

1. Support Vector Machines (SVM) with Hand-crafted Features: This approach emphasizes resource efficiency, operating effectively on a variety of common hardware setups, such as tablets and low-power devices. Diverse keypoint detectors and feature descriptors are employed, including AGAST, KAZE, SIFT, and VGG, to extract features from segmented solar cell images. The features are encoded via the Vectors of Locally Aggregated Descriptors (VLAD) methodology, and linear or radial basis function (RBF) SVMs are used for classification.
2. Convolutional Neural Networks (CNNs): Utilizing a transfer learning approach, the Vgg-19 architecture is fine-tuned to predict defect probabilities. This method operates on high GPU requirements but demonstrates higher classification performance. It involves data augmentation and a regression-based network to output continuous defect probabilities that are later categorized into four levels based on likelihood.

Results

The CNN approach achieved an average classification accuracy of 88.42%, outperforming the SVM's 82.44%. These methods make extensive use of a dataset comprising 2,624 solar cell images extracted from EL images of monocrystalline and polycrystalline PV modules. Monocrystalline module analysis revealed limited disparity between the CNN and SVM performances, whereas, for polycrystalline modules, the CNN displayed superior accuracy due to its enhanced capacity to manage varying textures.

A detailed analysis via t-distributed Stochastic Neighbor Embedding (t-SNE) demonstrated effective clustering of defect likelihoods in the CNN's feature space, facilitating accurate differentiation between defective and functional cells. Moreover, qualitative assessments endorsed the CNN’s accuracy, with Class Activation Maps (CAMs) elucidating discriminative regions of interest in cell inspections.

Implications and Future Outlook

The successful application of automated defect classification in PV cells can significantly streamline solar module inspections, reduce operational costs, and fortify the dependability of inspecting personnel-less solar power fields. By automating defect detection, operators can preemptively replace or repair modules that risk reducing the efficiency of solar power plants, maintaining optimal energy outputs.

Future work may involve scaling these approaches to handle larger datasets, integrating additional defect categories, or enhancing models to predict not just defect presence but specific defect types. Further developments could explore the incorporation of real-time defect tracking and large-scale deployment on operational solar farms.

By evolving the studied methods, there is potential to innovate a standardized, automated inspection protocol that can redefine maintenance strategies within photovoltaic energy production — paving the way for more sustainable and reliable energy solutions.