Source-linked AI summary
Solar Cell Surface Defect Inspection Based on Multispectral Convolutional Neural Network
Haiyong Chen, Yue Pang, Qidi Hu, Kun Liu
TL;DR
Solar-cell surface defects are difficult to detect because defect characteristics vary across shape, scale, and spectrum amid complex, random backgrounds. The paper selects a CNN structure and builds a multispectral CNN using separate spectral branches; experiments report higher accuracy and adaptability, while small linear defects remain challenging.
Problem
Solar-cell defects vary in shape, scale, and spectrum, while random lattice textures and non-uniform brightness make defect features difficult to distinguish.
Method
The paper selects a CNN structure by varying convolutional depth and kernel size, then sends separated red, green, and blue spectra through parallel CNNs whose features are fused for prediction.
Results
The multispectral solar-cell CNN improves detection of most defects and raises the detection rate of non-defective pictures by about 1.4%.
Takeaways & Limitations
The multispectral model has higher accuracy and stronger adaptability for solar-cell surface-defect detection, while remaining suitable for real-time processing.
Takeaways & Limitations
The model has weaker feature-extraction ability for small-area, linear defects such as broken gates and scratches.
Abstract
from arXiv · showhide
Similar and indeterminate defect detection of solar cell surface with heterogeneous texture and complex background is a challenge of solar cell manufacturing. The traditional manufacturing process relies on human eye detection which requires a large number of workers without a stable and good detection effect. In order to solve the problem, a visual defect detection method based on multi-spectral deep convolutional neural network (CNN) is designed in this paper. Firstly, a selected CNN model is established. By adjusting the depth and width of the model, the influence of model depth and kernel size on the recognition result is evaluated. The optimal CNN model structure is selected. Secondly, the light spectrum features of solar cell color image are analyzed. It is found that a variety of defects exhibited different distinguishable characteristics in different spectral bands. Thus, a multi-spectral CNN model is constructed to enhance the discrimination ability of the model to distinguish between complex texture background features and defect features. Finally, some experimental results and K-fold cross validation show that the multi-spectral deep CNN model can effectively detect the solar cell surface defects with higher accuracy and greater adaptability. The accuracy of defect recognition reaches 94.30%. Applying such an algorithm can increase the efficiency of solar cell manufacturing and make the manufacturing process smarter.
1 Introduction
Solar-cell surface defects threaten manufacturing quality and require timely detection. Polycrystalline cells are especially challenging because random lattice textures and varying defect colors create complex, non-uniform backgrounds.
- Surface defects including broken gates, paste spots, thick lines, dirty cells, missing corners, scratches, and color aberrations should be detected and eliminated during manufacturing.Timely removal helps avoid quality damage to solar-cell modules in subsequent production.
- Image-based machine vision can improve solar-cell module production quality and increase module lifetime.
- Polycrystalline silicon cells contain randomly shaped and positioned lattice particles that produce non-uniform, complex background textures.
- Random variation in defect color features makes visual inspection of multiple defects difficult against polycrystalline-cell backgrounds.
2 Related works on solar cell surface detection
Prior methods addressed surface defects across several material and imaging settings, but solar-cell inspection remains difficult because defects vary across spectra and complex polycrystalline backgrounds. The paper develops optimized CNN and multispectral approaches to automate feature extraction and inspection.
- Prior surface-inspection methods: Existing machine-vision methods classify defect inspection by texture type, including non-textured, repeated-pattern, homogeneous-textured, and non-homogeneous-textured surfaces.
- Solar-cell inspection gap: Visible-spectrum solar-cell defects have diverse appearances, while non-uniform textures and non-Gaussian color distributions weaken manually extracted defect features.
- Research problems: The paper identifies two central problems: highlighting multiple defect characteristics using multiple spectra and automatically extracting multispectral features for inspection.
- Proposed approach: A high-dynamic camera collects multispectral images, after which a defect dataset, optimized solar-cell CNN, and multispectral CNN are developed.
- CNN optimization: The CNN depth and convolution-kernel size are evaluated to establish an optimized model for distinguishing multiple defect features.
- Multispectral modeling: The multispectral model extracts multispectrum features, and cross-validation evaluates its comparison with the solar-cell CNN for robustness and resistance to overfitting.
- Stated novelty: The paper reports deep-learning-based solar-cell surface-defect inspection as a novel application in its stated research context.
3 Methods
The method analyzes spectral differences among polycrystalline-silicon solar-cell defects and builds CNN-based inspection models around these differences. It compares CNN structure choices, prepares a segmented dataset, and combines three spectral branches for multispectral prediction.
- 3.1 Multispectral defect feature analysis: Different defects show different spectral contrasts: broken gates, thick lines, and dirty cells are clearer in red or green, whereas scratches are stronger in green or blue.Paste spots and color differences remain evident across all three spectra, while lattice interference weakens blue-spectrum contrast for some defects.
- 3.1 Dataset construction: Sliding windows of 469×469 pixels with a 235-pixel stride produce 49 blocks per original image and expand the dataset to 15,330 defect-free and 5,915 defective images.The blocks are manually screened and classified into broken gates, paste spots, dirty cells, thick lines, scratches, and color differences.
- 3.1 Multispectral defect feature analysis: Polycrystalline silicon cells contain randomly distributed lattice particles and varying defect colors, creating non-uniform backgrounds that complicate multiple-defect inspection.The dataset includes six defect types whose shapes, scales, and spectral characteristics differ substantially.
- 3.2 Solar cell CNN model structure: The baseline CNN combines input, convolution, activation, pooling, and fully connected layers to extract image features and classify defects.Convolution extracts features, pooling reduces feature-map size, and the fully connected layer supplies the network output through a Softmax classifier.
- 3.2.1 Solar cell CNN model design: Model depth and convolution-kernel size are varied across candidate CNN structures, after which the best-performing structure is selected for multispectral experiments.The design is based on AlexNet and evaluates how depth and width changes affect detection performance.
- 3.3 Multispectral solar cell CNN model structure: The multispectral CNN separates the original image into three spectra, processes them through parallel convolutional feature-extraction branches, concatenates their outputs, and produces a final prediction.This structure is designed to exploit defect-specific spectral distinctions while separating complex background and defect features.
4 Experimental results analysis and discussion
Experiments evaluate CNN structure choices, multispectral feature extraction, cross-validation stability, and comparisons across defect tasks and baseline methods. The multispectral CNN generally improves defect discrimination and adaptability, while image segmentation and class imbalance constrain performance.
- CNN structure selection: A five-layer CNN improved precision by 2% over a three-layer model, while enlarging the kernel yielded precision and recall gains within 0.6%.The selected structure balanced performance against increased training time and pressure from deeper or wider models.
- Segmentation limitation: Too-small image segments caused overfitting because defective cells comprise only about 2% of total production.Smaller splits also reduce recognition speed and intensify defective-sample imbalance.
- Cross-validation comparison: Five-fold cross-validation showed higher recall and improved detection rates for dirty cells, thick lines, broken gates, color differences, and paste spots, while scratch detection fell about 0.5%.Non-defective-cell detection increased by 1.3%, and results showed no large fold-to-fold fluctuation.
- Multispectral feature extraction: Multispectral features separated defect and background texture more effectively, with red and green spectra especially informative for dirty-cell defects.Blue-spectrum features emphasized lattice and grating lines, whereas red and green features produced more accurate defect representations.
- 94.30% accuracy is reported for the multispectral deep CNN’s solar-cell defect recognition.
- Multi-class comparison: In multi-classification, multispectral CNN accuracy was 2 to 6 percentage points higher than the solar-cell CNN for each defect.Multi-class precision averaged 8% below binary classification, partly because some defect categories had few images and the dataset was unbalanced.
- Traditional-method comparison: LBP+HOG-SVM and Gabor-SVM had precision about 10% lower and recall about 8% lower than CNN methods on complex, textured cell surfaces.Manual texture features were less adaptable to random defect shapes and heterogeneous backgrounds containing lattice and grid-line patterns.
5 Conclusion
The multi-spectral solar cell CNN improves defect-background separation and overall detection adaptability, but remains weaker for small-area and linear defects. It maintains real-time processing while leaving model optimization and broader defect recognition for future work.
- The multi-spectral CNN extracts multiple spectral features, separates defect and background characteristics, and improves detection rates for most solar-cell defects.The reported detection rate for non-defective pictures increased by about 1.4%.
- Small-area and linear defects, including broken gates and scratches, have relatively low detection rates because their features are harder for the model to extract.By contrast, larger defects such as thick lines, dirty cells, paste spots, and color differences achieve higher detection rates.
- 3.66s is required to detect 100 images, only 0.59s more than the original CNN's 4.25s, while still meeting real-time detection needs.
- Future work will optimize recognition of linear and small-area defects, accelerate multithreaded training, and test recognition of more surface-defect types.Planned extensions also include displaying extracted features for accurate defect localization and using VGG or ResNet as an RGB-model base network to reduce training time and over-fitting.