Dust accumulation on solar panel surfaces is one of the primary causes of power loss in photovoltaic systems. The presence of dust particles with varying shapes, sizes, and chemical compositions significantly reduces the transmittance of solar radiation, thereby degrading the performance of solar panels. In regions with diverse climatic and geological conditions, the characteristics of dust differ greatly, making it essential to identify the type of dust on solar panels for optimal site selection and maintenance scheduling. Traditional methods such as scanning electron microscopy (SEM) are expensive and time-consuming. Therefore, developing an efficient and lightweight deep learning model for dust identification on solar panels is of great practical value. In this work, I propose an improved ShuffleNetV2 model that integrates Mish activation, mixed depthwise convolution, and coordinate attention mechanism to accurately classify dust particles on solar panels. The model achieves high accuracy while maintaining low computational complexity, making it suitable for deployment on portable or edge devices. Experimental results on a dataset collected from four different regions demonstrate that the proposed method outperforms several classical lightweight networks, including MobileNetV2, ResNet34, and GhostNet, in terms of both accuracy and model efficiency.
1. Introduction
With the global push toward carbon neutrality, solar energy has become a cornerstone of renewable energy systems. However, the efficiency of solar panels is highly sensitive to environmental factors, particularly dust deposition. Dust particles scatter and absorb incident sunlight, reducing the amount of light reaching the photovoltaic cells. Studies have shown that the shape and composition of dust particles vary by location, influencing the extent of performance degradation. For example, angular particles cause greater transmittance loss than spherical ones. To mitigate these effects, regular cleaning of solar panels is required, but identifying the type of dust is crucial for planning cleaning schedules and selecting appropriate cleaning methods. Recent advances in convolutional neural networks (CNNs) have enabled automated image-based classification of dust particles. However, due to the limited computational resources of mobile inspection devices, lightweight models are preferred. ShuffleNetV2 is a state-of-the-art efficient CNN architecture, but its standard design may not be optimal for dust recognition tasks. In this paper, I enhance ShuffleNetV2 by introducing three key modifications: (1) replacing ReLU with Mish activation to allow better gradient flow, (2) employing mixed depthwise convolutions with kernels of sizes 3×3, 5×5, and 7×7 to capture multi-scale features, and (3) inserting coordinate attention (CA) modules to focus on discriminative spatial regions while reducing parameters. The proposed model significantly improves classification accuracy for solar panel dust images while maintaining a low computational footprint.
2. Related Work
Numerous researchers have investigated the physical and chemical properties of dust on solar panels. Mishra et al. (2015) classified dust particles into six shapes based on SEM images. Kazem et al. (2016) studied dust from six cities in Oman and found that particle morphology affects solar panel performance. Tanesab et al. (2019) compared dust from two Australian locations and reported that angular particles caused higher power degradation than spherical ones. Wu et al. (2021) predicted relative transmittance loss using cubic and spherical dust models. These studies highlight the importance of dust shape identification for solar panel management. On the deep learning front, CNNs have been applied to particle classification. Yin et al. (2020) combined CNNs with SEM for atmospheric particle classification. Maffezzoli et al. (2022) used deep neural networks for ice core particle identification. Ayyagari et al. (2022) proposed a CNN-LSTM model for industrial dust classification. However, these approaches often rely on heavy networks not suitable for resource-constrained devices. Lightweight architectures like ShuffleNetV2 are promising but need adaptation for dust-specific features. My work bridges this gap by tailoring ShuffleNetV2 to solar panel dust recognition.
3. Baseline ShuffleNetV2 Model
ShuffleNetV2 is a lightweight CNN designed for mobile devices. Its basic unit consists of a channel split operation followed by two branches. In the stride-1 unit, the right branch employs pointwise convolutions (1×1) for channel fusion, a 3×3 depthwise convolution for spatial feature extraction, and another pointwise convolution. The left branch is an identity mapping. The outputs are concatenated and channel-shuffled to promote cross-channel information exchange. In the stride-2 unit, both branches use depthwise convolutions with stride 2 for downsampling. The design follows four practical guidelines: using equal channel widths, avoiding excessive group convolution, minimizing network fragmentation, and reducing element-wise operations. The original model uses ReLU activation. The architecture is efficient, but its performance on dust images can be improved by addressing the limitations of ReLU, the fixed kernel size, and the lack of spatial attention.
4. Proposed Improvements
4.1 Mish Activation Function
ReLU is widely used but suffers from the “dying ReLU” problem where negative inputs yield zero gradients, stalling learning. Mish is a smooth, non-monotonic activation defined as:
$$ f(x) = x \cdot \tanh(\ln(1+e^x)) $$
Mish allows small negative gradients to flow through, enabling deeper propagation of information. It is self-regularizing and often leads to better generalization. I replace all ReLU activations in ShuffleNetV2 with Mish. This change does not increase parameters but improves accuracy by 0.65 percentage points on our dataset, as shown in Table 1.
4.2 Mixed Depthwise Convolution
Dust particles on solar panels exhibit diverse sizes. A single 3×3 kernel may not capture all scales effectively. Using multiple kernel sizes as in Inception increases parameters significantly. Mixed depthwise convolution (MixConv) addresses this by applying different kernel sizes to different channel groups. I split the input channels into three groups and apply 3×3, 5×5, and 7×7 depthwise convolutions respectively, then concatenate the outputs and use a 1×1 pointwise convolution to fuse them. The structure is shown in Figure 4 of the original paper. This allows the model to learn multi-scale features with minimal extra cost. However, adding larger kernels increases FLOPs and parameters. I tested combinations: (3,5), (3,5,7), and (3,5,7,9). The best trade-off was achieved with (3,5,7) which improved accuracy by 1.61% over baseline while increasing FLOPs by only 0.05×10^8.
4.3 Coordinate Attention Mechanism
Standard channel attention (SENet) loses spatial information. CBAM uses 7×7 convolution for spatial attention but cannot capture long-range dependencies. Coordinate attention (CA) encodes both channel and spatial information by performing global average pooling along horizontal and vertical directions separately, then concatenating and applying a shared convolution, batch norm, and nonlinearity, followed by two separate sigmoid activations. The output is element-wise multiplied with the input. CA has negligible computational overhead. In ShuffleNetV2, the pointwise convolution at the tail of the right branch accounts for a large portion of computation because it operates on the full channel dimension. Replacing this 1×1 convolution with CA not only reduces parameters (since CA has fewer parameters than a full pointwise convolution) but also enhances feature representation by emphasizing informative spatial positions. I replaced the final pointwise convolution in both stride-1 and stride-2 basic units with CA. This reduced FLOPs from 1.50×10^8 to 1.08×10^8 and parameters from 2,278,604 to 1,992,700, while boosting accuracy by 1.54% (see Table 1).
5. Experimental Setup
5.1 Dataset
I collected 239 raw microscope images of dust from four regions in China: Yijinhuoluo Banner (61 images), Yulin (65), Xi’an (44), and Weinan (69). Dust particles in these images show distinct morphologies: Yijinhuoluo has flocculent particles, Yulin has spherical-like particles, etc. To prevent overfitting, I applied offline augmentation: brightness adjustment, horizontal flip, rotation, and translation, yielding 718 images (170 from Yijinhuoluo, 192 from Yulin, 152 from Xi’an, 204 from Weinan). During training, online augmentation (random crop to 224×224 and normalization) was applied. The dataset was split into 80% training and 20% testing.
5.2 Training Details
All experiments were implemented in PyTorch 1.71. Initial learning rate: 0.001. Epochs: 100. Batch size: 24. Optimizer: Adam. Loss: cross-entropy. Weights initialized with Gaussian (mean=0, std=0.01). Transfer learning was used by loading pre-trained ShuffleNetV2 weights. Evaluation metrics: accuracy (Acc), number of parameters (Params), and floating-point operations (FLOPs). Formulas for a convolutional layer:
$$ \text{Params}_{conv} = (K_h \times K_w \times C_{in} + 1) \times C_{out} $$
$$ \text{FLOPs}_{conv} = [(K_h \times K_w \times C_{in} + 1) \times C_{out}] \times H_{out} \times W_{out} $$
6. Results and Discussion
6.1 Ablation Studies
Table 1 summarizes the ablation experiments. Starting from the baseline ShuffleNetV2 (Acc: 87.44%, Params: 2,278,604, FLOPs: 1.50×10^8), each improvement contributed positively. Adding Mish alone increased accuracy by 0.65% with identical parameters. Adding CA alone reduced parameters to 1,992,700 and FLOPs to 1.08×10^8 while improving accuracy by 1.54%. Combining Mish and CA yielded 90.12% accuracy with FLOPs of 1.07×10^8. Then, mixing depthwise convolutions (3,5) with Mish+CA gave 91.28% but slightly increased parameters and FLOPs. The best result was obtained with (3,5,7) kernels, Mish, and CA: accuracy 92.25%, FLOPs 1.14×10^8, and Params 2,031,288. Using larger kernels (3,5,7,9) degraded accuracy to 91.19% due to overfitting or excessive parameters.
| Model Variant | Accuracy (%) | Params (count) | FLOPs (×10^8) |
|---|---|---|---|
| ShuffleNetV2 (baseline) | 87.44 | 2,278,604 | 1.50 |
| + Mish | 88.09 | 2,278,604 | 1.49 |
| + CA | 88.98 | 1,992,700 | 1.08 |
| + Mish + CA | 90.12 | 1,992,700 | 1.07 |
| + (3,5) MixConv + Mish + CA | 91.28 | 2,009,404 | 1.10 |
| + (3,5,7) MixConv + Mish + CA | 92.25 | 2,031,288 | 1.14 |
| + (3,5,7,9) MixConv + Mish + CA | 91.19 | 2,059,260 | 1.19 |
6.2 Comparison with Other Lightweight Models
To validate the effectiveness, I compared the proposed improved ShuffleNetV2 (MixConv (3,5,7) + Mish + CA) with MobileNetV2, ResNet34, and GhostNet under identical conditions. Results are in Table 2. My method achieved the best accuracy (92.25%) while having the lowest FLOPs (1.14×10^8) and competitive parameters. MobileNetV2 reached 91.14% accuracy but required 3.20×10^8 FLOPs. ResNet34 achieved 91.60% accuracy but had 18.21×10^8 FLOPs, which is 16 times higher. GhostNet had 87.74% accuracy with 1.50×10^8 FLOPs. Thus, my model offers an excellent balance between accuracy and computational cost, making it highly suitable for real-time solar panel dust identification on edge devices.
| Model | Accuracy (%) | Params (count) | FLOPs (×10^8) |
|---|---|---|---|
| ShuffleNetV2 (baseline) | 87.44 | 2,278,604 | 1.50 |
| Improved ShuffleNetV2 (Ours) | 92.25 | 2,031,288 | 1.14 |
| MobileNetV2 | 91.14 | 3,504,872 | 3.20 |
| ResNet34 | 91.60 | 11,689,512 | 18.21 |
| GhostNet | 87.74 | 5,183,016 | 1.50 |
6.3 Effect of Data Augmentation
To show the importance of data augmentation, I trained the baseline ShuffleNetV2 on the original 239 images (without augmentation) and on the augmented 718 images. Accuracy increased from 86.93% to 87.44% (+0.51%), demonstrating that augmentation mitigates overfitting and improves generalization for solar panel dust classification.
6.4 Effect of Activation Function
Figure 8 in the original paper (not reproduced here) compared ReLU and Mish training curves on baseline ShuffleNetV2. Mish yielded faster convergence and higher final accuracy (88.09% vs 87.44%). This confirms that Mish’s smooth gradient flow benefits dust image feature learning.
6.5 Effect of Coordinate Attention
Integrating CA into the baseline model (without other changes) increased accuracy from 87.44% to 88.98% while reducing both parameters and FLOPs. CA excels at encoding spatial location information, which is crucial for discriminating between different dust morphologies that often occupy specific regions in the microscope images.
7. Conclusion
In this work, I proposed an improved ShuffleNetV2 model for identifying dust on solar panels. By replacing ReLU with Mish, introducing mixed depthwise convolutions (3×3, 5×5, 7×7), and substituting the tail pointwise convolution with coordinate attention, the model achieves 92.25% accuracy on a four-class dust dataset, while maintaining extremely low computational cost (1.14×10^8 FLOPs). Compared with MobileNetV2, ResNet34, and GhostNet, my model offers superior accuracy-to-complexity trade-off. The lightweight design makes it ideal for integration into portable dust inspection devices for solar panel maintenance. Future work will focus on establishing a direct correlation between identified dust types and the power output degradation of solar panels, enabling predictive cleaning schedules. Additionally, expanding the dataset with more regional dust samples will further enhance the model’s robustness.