In recent years, the increasing global energy demand and environmental concerns have accelerated the adoption of renewable energy sources, with photovoltaic technology playing a pivotal role. Solar panels, particularly polycrystalline silicon-based ones, are widely used in photovoltaic systems due to their cost-effectiveness and efficiency. However, defects such as cracks, black spots, and broken grids often occur during manufacturing or operational phases, significantly reducing the energy conversion efficiency and lifespan of solar panels. Traditional defect classification methods, including physical inspections and laser scanning, are time-consuming, prone to human error, and may cause secondary damage. Meanwhile, conventional machine learning approaches require large datasets and struggle with generalization in complex industrial environments. To address these limitations, we propose an enhanced prototypical network method for few-shot learning, which improves feature embedding and similarity measurement for accurate defect classification in photovoltaic systems.
The growing reliance on solar energy underscores the importance of maintaining the integrity of photovoltaic modules. Defects in solar panels can lead to substantial power losses and safety hazards, making automated classification systems essential. Existing methods, as summarized in Table 1, highlight the trade-offs between accuracy and efficiency. For instance, manual techniques like acoustic wave analysis offer high precision but are impractical for large-scale deployments. In contrast, machine learning models such as support vector machines (SVM) and deep neural networks (DBNs) provide automation but often overfit with limited data. Few-shot learning, particularly prototypical networks, emerges as a promising solution by leveraging small datasets to generalize across tasks. Our work builds on this foundation by integrating advanced neural architectures and optimized metrics tailored for photovoltaic applications.
| Method Category | Approach Overview | Advantages | Limitations |
|---|---|---|---|
| Manual Physical Methods | Apply loads, lasers, or ultrasound to measure voltage/current | High accuracy, mature technology | Time-consuming, subjective, risk of damage |
| Machine Learning Methods | Use algorithms like SVM or CNN to build classification models | Automated, cost-effective, stable | Data-dependent, prone to overfitting |
Prototypical networks, introduced by Snell et al., utilize a metric-based approach to classify samples by computing prototypes for each class. The prototype for a class \(k\) is defined as the mean of embedded support samples:
$$c_k = \frac{1}{|S_k|} \sum_{(x_i, y_i) \in S_k} f_\phi(x_i)$$
where \(S_k\) represents the support set for class \(k\), \(x_i\) denotes a sample, \(y_i\) is its label, and \(f_\phi\) is the embedding function parameterized by \(\phi\). This method maps samples into a feature space where classification is performed by measuring distances to prototypes. While effective for simple images, the standard prototypical network struggles with complex photovoltaic defect samples due to limited feature extraction capabilities and suboptimal similarity metrics.
To overcome these challenges, we introduce three key improvements. First, we replace the conventional ConvNet4 backbone with AResNet, a ResNet18 variant enhanced with dual attention mechanisms (spatial and channel). This architecture deepens the network while preserving feature integrity through residual blocks. The attention modules emphasize relevant features, such as defect patterns in solar panels, while suppressing noise. The residual learning in AResNet addresses gradient issues in deep networks, formulated as:
$$y = F(x, \{W_i\}) + x$$
where \(x\) is the input, \(F\) represents the residual mapping, and \(W_i\) denotes weights. This allows the network to learn incremental features, crucial for capturing subtle defects in photovoltaic modules.
Second, we modify the pre-training strategy by incorporating a mix of external auxiliary data (e.g., PVELAD dataset) and task-specific photovoltaic defect samples. This semi-supervised approach aligns the model with the target data distribution, enhancing its applicability to real-world solar panel inspections. The pre-training minimizes the loss function:
$$\mathcal{L} = -\sum_{i=1}^{N} \log \frac{\exp(-d(f_\phi(x_i), c_{y_i}))}{\sum_{j=1}^{N} \exp(-d(f_\phi(x_i), c_j))}$$
where \(d\) is the distance metric, and \(N\) is the number of classes. By including photovoltaic samples early, the model prioritizes task-relevant features, improving generalization.
Third, we adopt squared Euclidean distance as the similarity metric instead of Euclidean or cosine distance. The squared Euclidean distance between two samples \(X\) and \(Y\) in an \(n\)-dimensional space is computed as:
$$d(X,Y) = \sum_{i=1}^{n} (x_i – y_i)^2$$
This metric amplifies inter-class differences, making it easier to distinguish defects like micro-cracks or discoloration in solar panels. Compared to Euclidean distance, it eliminates the square root operation, reducing computational overhead and enhancing efficiency for photovoltaic applications.
We evaluated our method on a customized photovoltaic defect dataset comprising 20 classes, including shape-based defects (e.g., scratches, cracks), brightness anomalies (e.g., hotspots, mismatches), and positional issues (e.g., broken grids, dirt). The dataset combines public sources and real-time captures, with images preprocessed to remove noise and standardize dimensions. Experiments were conducted under few-shot settings, such as 4-way and 6-way classification, with varying shot numbers (e.g., 10-shot, 30-shot). The results, summarized in Table 2, demonstrate that our improved prototypical network achieves higher accuracy and reduced time compared to baseline methods like K-nearest neighbors (KNN) and standard prototypical networks.
| Model | Scenario | Task | Accuracy (%) | Time (min) |
|---|---|---|---|---|
| KNN | General | 4-class | 29.1 | 178 |
| Standard PN | Photovoltaic | 4-way 10-shot | 68.2 | 114 |
| Improved PN | Photovoltaic | 4-way 10-shot | 70.3 | 127 |
| Improved PN | Photovoltaic | 6-way 30-shot | 72.3 | 135 |
Ablation studies further validate the contributions of each improvement. As shown in Table 3, using AResNet alone boosts accuracy by approximately 5% over ConvNet4, while the combined enhancements yield the best results. The squared Euclidean distance consistently outperforms other metrics, emphasizing its suitability for photovoltaic defect differentiation. These findings highlight the importance of tailored architectures and metrics in few-shot learning for solar panel inspections.
| Backbone | Pre-training Data | Distance Metric | Accuracy (%) | Time (min) |
|---|---|---|---|---|
| ConvNet4 | Auxiliary only | Euclidean | 61.8 | 74 |
| AResNet | Auxiliary + Photovoltaic | Squared Euclidean | 71.7 | 146 |
| AResNet | Auxiliary only | Cosine | 65.6 | 122 |
The integration of attention mechanisms in AResNet enables focused feature extraction from complex photovoltaic samples. For example, spatial attention weights salient regions like crack edges, while channel attention emphasizes spectral features indicative of defects. The residual blocks facilitate training deeper networks without degradation, crucial for handling diverse solar panel conditions. The pre-training strategy ensures that the model captures both general and specific features, reducing the domain gap between auxiliary and target data. This is particularly important for photovoltaic systems, where environmental factors like lighting and angle variations can affect image quality.
In terms of computational efficiency, the squared Euclidean distance simplifies the similarity calculation by avoiding square roots, which is beneficial for real-time applications in photovoltaic monitoring. The distance between a query sample \(q\) and a prototype \(c_k\) is computed as:
$$d(q, c_k) = \sum_{i=1}^{m} (q_i – c_{k,i})^2$$
where \(m\) is the feature dimension. This linear operation accelerates inference, making our method feasible for embedded systems in solar farms.
Future work could explore multi-modal data fusion, such as combiningelectroluminescence and thermal images of solar panels, to enhance defect detection. Additionally, adaptive metric learning could further improve the prototypical network’s robustness to novel defects in photovoltaic modules. The scalability of this approach makes it suitable for large-scale solar energy installations, where automated inspection is critical for maintenance and efficiency.
In conclusion, our improved prototypical network addresses the limitations of traditional methods by leveraging advanced neural architectures, optimized training strategies, and enhanced similarity metrics. Experimental results on photovoltaic defect datasets confirm significant gains in accuracy and efficiency, demonstrating the potential of few-shot learning for sustainable energy applications. As the demand for solar panels grows, such innovations will play a vital role in ensuring the reliability and performance of photovoltaic systems worldwide.