With the increasing adoption of solar panels in renewable energy systems, automated inspection of surface damage and contamination has become critical for maintaining energy conversion efficiency. This paper presents an optimized FasterNet architecture integrated with parameter-free attention mechanisms for intelligent classification of six common solar panel surface conditions: bird droppings, dust accumulation, electrical burns, physical damage, snow coverage, and clean surfaces.
Methodology
The proposed framework combines enhanced feature extraction with efficient computation through three key innovations:
1. Data Preparation and Augmentation
A dataset containing 4,320 images (640×640 pixels) of solar panel surfaces was created through field collection and web crawling, with distribution as shown in Table 1. Spatial transformations including random rotation (±15°), horizontal flipping, and translation (±10% offset) were applied for data augmentation.
| Class | Original | Augmented |
|---|---|---|
| Bird Droppings | 347 | 720 |
| Clean Surface | 401 | 720 |
| Dust Accumulation | 512 | 720 |
| Electrical Burns | 289 | 720 |
| Physical Damage | 326 | 720 |
| Snow Coverage | 417 | 720 |
2. Enhanced FasterNet Architecture
The baseline FasterNet structure was modified by inserting SimAM attention modules after the first and third PConv blocks. The energy function for SimAM attention is defined as:
$$e_k^* = \frac{4(\hat{\sigma}^2 + \lambda)}{(t_k – \hat{\mu})^2 + 2\hat{\sigma}^2 + 2\lambda}$$
where $\hat{\mu}$ and $\hat{\sigma}^2$ represent channel-wise mean and variance, respectively. The attention-weighted feature map $\tilde{X}$ is calculated as:
$$\tilde{X} = \text{sigmoid}\left(\frac{1}{E}\right) \odot X$$
3. Optimization Strategy
The Adam optimizer with adaptive learning rate was employed with initial parameters:
$$m_t = \beta_1 m_{t-1} + (1-\beta_1)g_t$$
$$v_t = \beta_2 v_{t-1} + (1-\beta_2)g_t^2$$
$$\theta_t = \theta_{t-1} – \alpha \frac{m_t}{\sqrt{v_t} + \epsilon}$$
where $\beta_1=0.9$, $\beta_2=0.999$, and $\alpha=3\times10^{-4}$.
Experimental Results
The enhanced solar panel inspection system achieved superior performance compared to conventional models:
| Model | Accuracy (%) | F1-Score (%) | Params (MB) |
|---|---|---|---|
| ResNet-50 | 72.47 | 73.35 | 81.3 |
| VGG-16 | 69.58 | 71.59 | 512 |
| MobileNet-small | 76.84 | 77.48 | 4.9 |
| Proposed | 78.28 | 77.45 | 4.0 |
The attention mechanism improved feature discrimination for challenging solar panel conditions, particularly between dust accumulation and clean surfaces. The model’s computational efficiency enables real-time deployment with:
$$ \text{Throughput} = \frac{1280\ \text{FPS}}{\text{Batch Size}} \quad \text{on RTX 3060 GPU}$$
Conclusion
This enhanced FasterNet architecture demonstrates effective solar panel surface inspection capabilities through:
- Parameter-free attention mechanism for improved feature discrimination
- Lightweight design suitable for edge deployment
- Adaptive optimization for stable convergence
The system achieves 78.28% accuracy in classifying six common solar panel surface conditions while maintaining 4MB model size, providing a practical solution for automated photovoltaic array maintenance. Future work will focus on multi-spectral analysis for early-stage defect detection in solar panel systems.