Solar energy has become a central pillar in the global transition toward clean and renewable power sources. As the installed capacity of photovoltaic systems continues to expand, maintaining the operational efficiency and safety of solar panels has become a critical engineering challenge. In my research, I focus on the surface defect detection of polycrystalline silicon solar panels, which are widely deployed in both centralized power plants and distributed rooftop systems. These panels inevitably suffer from various types of defects during manufacturing and long-term outdoor operation. Defects such as cracks, broken grids, black cores, dust accumulation, bird droppings, and physical damage can significantly reduce the energy conversion efficiency of solar panels and, in severe cases, pose safety hazards. Therefore, developing an accurate and efficient automated detection algorithm is of great practical significance.
Traditional manual inspection methods are not only time-consuming but also highly subjective, as their accuracy heavily depends on the experience and skill level of the inspectors. To overcome these limitations, I have developed deep learning-based approaches for detecting surface defects on solar panels in two distinct application scenarios: production-line inspection using electroluminescence (EL) imaging and outdoor maintenance inspection using visible-light imaging. This article presents the complete methodology, the model architectures I designed, the experimental validation, and the practical detection system I implemented.
Overview of the Proposed Approach
My work addresses two complementary detection scenarios. In the production-line scenario, solar panels are inspected using an EL imaging technique, which can reveal internal structural defects such as micro-cracks and broken fingers. However, EL images typically have low contrast, complex backgrounds, and tiny defect regions. To address these challenges, I proposed a detection model named EVT-Net, based on an efficient vision Transformer backbone. In the outdoor operation scenario, solar panels are photographed under natural illumination using visible-light cameras. These images present different challenges, including variable defect shapes, inconsistent scales, and low contrast between defects and the panel background. For this scenario, I designed MCAM-Net, an improved YOLO11-based model that integrates a novel attention mechanism, multi-scale convolution modules, and deformable convolution.
Related Work and Theoretical Foundations
Defect detection techniques for solar panels have evolved through three major stages. In the early stage, manual visual inspection dominated, relying entirely on human expertise. The second stage introduced traditional machine vision methods based on image processing algorithms, including threshold segmentation, edge detection, texture feature analysis, and transform-domain techniques. While these methods improved automation, they struggled with complex backgrounds, low contrast, and the need for manual feature design. The third and current stage leverages deep learning, which automatically learns hierarchical features from data and has substantially improved detection performance.
Convolutional neural networks (CNNs) form the backbone of most modern object detection systems. The basic convolution operation can be expressed as:
$$
Y[i,j,c] = \sum_{m=0}^{k_h-1} \sum_{n=0}^{k_w-1} \sum_{d=0}^{C_{in}-1} X[i+m, j+n, d] \cdot K[m, n, d, c]
$$
where $Y$ is the output feature map, $X$ is the input feature map, $K$ is the convolutional kernel, and $k_h$, $k_w$ are the kernel height and width, respectively. Pooling layers downsample feature maps to reduce computational load and increase translation invariance.
Object detection algorithms are generally categorized into two-stage detectors such as Faster R-CNN, and one-stage detectors such as the YOLO family and SSD. Two-stage detectors first generate region proposals and then classify each proposal, achieving high accuracy at the cost of speed. One-stage detectors directly regress bounding boxes and class probabilities in a single forward pass, offering better speed-accuracy trade-offs for real-time applications. In recent years, Vision Transformer models have also been introduced to object detection. These models use multi-head self-attention to capture global contextual relationships, providing a different inductive bias compared to CNNs.
To evaluate the performance of my models, I used standard metrics including precision (P), recall (R), mean average precision (mAP50 and mAP50-95), parameter count, and computational complexity measured in FLOPs. These metrics provide a comprehensive assessment of detection accuracy and computational efficiency.
EVT-Net for Electroluminescence Defect Detection
The production-line inspection of solar panels relies heavily on EL imaging. In my work, I used the EL2021 dataset, which is publicly available on the Roboflow platform. This dataset comprises 5,016 EL images of polycrystalline silicon solar panels, annotated with six defect types: black core, crack, broken finger, horizontal dislocation, short circuit, and thick line. During preliminary analysis, I discovered labeling issues in some samples, including misplaced bounding boxes and missing small defects. I therefore re-annotated problematic images using the LabelImg tool to create a more reliable dataset for model training and evaluation.
The challenges in EL image defect detection are threefold. First, defects have low contrast relative to the bright background of the solar panel. Second, the background texture of polycrystalline panels is rich and complex due to grain boundaries. Third, many defects are extremely small, making them difficult to distinguish from background variations. To address these challenges, I proposed the EVT-Net model, which is built upon the CAS-ViT architecture.
M-CASViT Backbone Network
CAS-ViT is an efficient vision Transformer designed for mobile applications. Its core innovation is the convolutional additive token mixer (CATM), which replaces the quadratic-complexity self-attention mechanism with a linear-complexity additive formulation. The similarity function of the original self-attention mechanism is:
$$
\text{Sim}(Q, K) = \exp(QK^T / \sqrt{d})
$$
where $Q$ and $K$ are the query and key matrices, and $d$ is the scaling factor. The computational complexity of this operation is $O(N^2)$, where $N$ is the number of tokens. In contrast, CATM uses the following formulation:
$$
\text{Sim}(Q, K) = \Phi(Q) + \Phi(K)
$$
$$
O = \Gamma(\Phi(Q) + \Phi(K)) \cdot V
$$
where $\Phi(\cdot)$ is a Sigmoid-based context mapping function and $\Gamma(\cdot)$ is an activation function. This reduces the complexity to $O(N)$, effectively eliminating matrix multiplication and Softmax operations while maintaining global modeling capability.
In my M-CASViT backbone, I made two key modifications to the original CAS-ViT architecture. First, I optimized the stage layer configuration from [2,2,4,2] to [1,1,2,1]. This adjustment was motivated by the observation that the computational cost of a Transformer layer increases quadratically with feature map resolution and linearly with depth. The relationship between layer configuration and model complexity is shown in Table 1 below.
| Stage Layers | Params (M) | FLOPs (G) | mAP50 (%) | mAP50-95 (%) |
|---|---|---|---|---|
| [2,2,4,2] | 8.19 | 28.98 | 92.72 | 66.69 |
| [2,2,2,2] | 7.40 | 26.78 | 92.63 | 65.71 |
| [1,1,1,1] | 5.87 | 21.29 | 91.62 | 63.84 |
| [1,1,2,2] | 7.26 | 23.09 | 92.76 | 66.75 |
| [1,1,2,1] | 6.26 | 22.39 | 92.75 | 66.73 |
As shown in Table 1, the configuration [1,1,2,1] achieves the best trade-off between detection accuracy and computational efficiency, reducing parameters by 1.93M and FLOPs by 6.59G compared to the original configuration, while simultaneously improving mAP50 from 92.72% to 92.75% and mAP50-95 from 66.69% to 66.73%.
Second, I designed a multi-branch stem module (M-Stem) to replace the standard convolutional downsampling used in the original CAS-ViT. The M-Stem processes the input through three parallel branches: a standard 3×3 convolution branch, an average pooling branch, and a max pooling branch. These branches are then concatenated and fused through a 1×1 convolution. The forward propagation can be expressed as:
$$
Y = f_{\text{conv}}^{3 \times 3}(X)
$$
$$
B_{\text{conv}} = f_{\text{conv}}^{3 \times 3}(Y), \quad B_{\text{avg}} = P_{\text{avg}}(Y), \quad B_{\text{max}} = P_{\text{max}}(Y)
$$
$$
Z = f_{\text{conv}}^{1 \times 1}(\text{Concat}(B_{\text{conv}}, B_{\text{avg}}, B_{\text{max}}))
$$
This multi-path design preserves detailed local features while also capturing global contextual information. The max pooling branch emphasizes high-response features, while the average pooling branch preserves background context, resulting in richer and more robust feature representations. The effectiveness of M-Stem was validated through a comparative experiment, with results summarized in Table 2.
| Module | P (%) | R (%) | Params (M) | FLOPs (G) | mAP50 (%) | mAP50-95 (%) |
|---|---|---|---|---|---|---|
| Standard Stem | 91.42 | 91.13 | 6.27 | 22.54 | 91.73 | 66.25 |
| M-Stem | 92.37 | 93.27 | 6.26 | 22.39 | 92.75 | 66.73 |
DDC-FPN for Multi-Scale Feature Fusion
To improve the detection performance for small-scale defects in EL images, I designed a dual-path dynamic calibration feature pyramid network (DDC-FPN). The standard FPN fuses multi-scale features with fixed weights, which limits adaptability. In contrast, DDC-FPN uses a dynamic dual-path calibration block (DDC Block) that processes features through two parallel paths: a channel path and a spatial path.
The channel path computes attention weights through global average pooling followed by two fully connected layers:
$$
A_c = \sigma(W_2 \cdot \delta(W_1 \cdot \text{GAP}(X)))
$$
The spatial path extracts spatial context using a sequence of lightweight convolutions:
$$
A_s = \sigma(C_{1 \times 1}^{(2)}(\text{GN}(\text{SiLU}(C_{1 \times 1}^{(1)}(X)))))
$$
A dynamic gating unit balances the contributions of the two paths:
$$
G = \sigma(W_2 \cdot \text{ReLU}(W_1 \cdot \text{GAP}(X)))
$$
$$
X_{\text{out}} = X + G \odot A_c + (1 – G) \odot A_s
$$
Table 3 compares the detection performance of EVT-Net with FPN versus DDC-FPN.
| Module | P (%) | R (%) | Params (M) | FLOPs (G) | mAP50 (%) | mAP50-95 (%) |
|---|---|---|---|---|---|---|
| FPN | 92.11 | 91.89 | 6.19 | 22.39 | 91.19 | 65.21 |
| DDC-FPN | 92.37 | 93.27 | 6.26 | 22.39 | 92.75 | 66.73 |
DDC-FPN improves mAP50 by 1.56% and mAP50-95 by 1.52% with a negligible parameter increase, demonstrating its effectiveness in multi-scale feature enhancement.
Decoupled Detection Head
The coupled detection head used in many single-stage detectors shares a single feature representation for both classification and regression tasks. This design can lead to gradient conflicts and suboptimal convergence. I introduced the Decoupled Head structure, which splits classification and regression into independent branches. The total loss function is defined as:
$$
L = \lambda_{\text{reg}} L_{\text{reg}} + L_{\text{obj}} + L_{\text{cls}} + \mathbb{1}_{l_1} L_{l_1}
$$
The regression loss uses IoU-based loss for better scale invariance:
$$
L_{\text{reg}} = \frac{1}{N_{\text{pos}}} \sum_{i \in \Omega_{\text{pos}}} \left[1 – \text{IoU}(B_{\text{pred}}^i, B_{\text{gt}}^i)\right]
$$
The objectness loss employs binary cross-entropy:
$$
L_{\text{obj}} = \frac{1}{N_{\text{all}}} \sum_{i=1}^{N_{\text{all}}} \left[o_i \log(\sigma(o_i^*)) + (1 – o_i) \log(1 – \sigma(o_i^*))\right]
$$
The classification loss is also binary cross-entropy, but applied only to the class branches of positive samples. Table 4 shows the comparison between coupled and decoupled heads.
| Head Type | P (%) | R (%) | Params (M) | FLOPs (G) | mAP50 (%) | mAP50-95 (%) |
|---|---|---|---|---|---|---|
| Coupled Head | 91.51 | 92.43 | 6.26 | 22.39 | 90.35 | 64.17 |
| Decoupled Head | 92.37 | 93.27 | 6.26 | 22.39 | 92.75 | 66.73 |
Experimental Validation of EVT-Net
I conducted a comprehensive comparison of EVT-Net against several state-of-the-art detection models, including Faster R-CNN, FCOS, RetinaNet, RT-DETR, and the YOLOv8 to YOLOv12 series. All experiments were performed on the same hardware platform with an NVIDIA RTX 4090 GPU, using identical training configurations to ensure fairness. The results are presented in Table 5.
| Model | P (%) | R (%) | Params (M) | FLOPs (G) | mAP50 (%) | mAP50-95 (%) |
|---|---|---|---|---|---|---|
| Faster R-CNN | 92.32 | 91.95 | 41.25 | 91.03 | 91.83 | 66.41 |
| FCOS | 89.71 | 88.06 | 31.85 | 78.73 | 88.92 | 63.87 |
| RetinaNet | 87.95 | 89.47 | 36.21 | 82.57 | 87.21 | 61.83 |
| RT-DETR | 88.24 | 90.33 | 19.87 | 56.90 | 89.75 | 64.56 |
| YOLOv8n | 90.56 | 91.71 | 3.01 | 8.20 | 90.83 | 65.84 |
| YOLOv9t | 92.89 | 90.02 | 2.01 | 7.90 | 91.15 | 66.21 |
| YOLOv10n | 92.27 | 90.45 | 2.71 | 8.40 | 91.58 | 66.75 |
| YOLO11n | 90.95 | 91.12 | 2.59 | 6.40 | 91.97 | 65.02 |
| YOLOv12n | 91.31 | 92.64 | 2.60 | 6.50 | 90.67 | 65.59 |
| EVT-Net | 92.37 | 93.27 | 6.26 | 22.39 | 92.75 | 66.73 |
EVT-Net achieved the highest mAP50 of 92.75% and mAP50-95 of 66.73% among all compared models. While its parameter count of 6.26M is higher than that of the YOLO series, it remains suitable for real-time inspection applications. The recall value of 93.27% indicates that EVT-Net effectively reduces missed detections, which is critical for quality control in solar panel manufacturing.
MCAM-Net for Visible-Light Defect Detection
In the outdoor operation and maintenance scenario, solar panels are inspected using visible-light images that capture defects such as dust, bird droppings, physical damage, snow, and electrical damage. I constructed the MS2025 dataset, combining images collected from a distributed photovoltaic station at a university campus with open-source data. The dataset contains 3,956 images spanning six categories including a non-defective category. All images were annotated using the LabelImg tool. The sample distribution is summarized in the experimental section of my thesis.
The outdoor visible-light detection task presents several unique challenges. Defects exhibit highly variable shapes and sizes. Some defects, particularly dust accumulation and early-stage damage, have low contrast against textured solar panel surfaces. Additionally, outdoor lighting conditions vary dramatically across different times of day and weather conditions. To address these challenges, I designed the MCAM-Net model based on YOLO11, incorporating three major improvements.
CPMSSA Attention Mechanism
The first improvement is a channel-prioritized multi-scale subspace attention mechanism (CPMSSA). This mechanism combines channel attention with a novel multi-scale subspace attention module to enhance the model’s sensitivity to low-contrast defects. The overall attention computation is:
$$
X_c = \text{CA}(X) \odot X
$$
$$
Y = \text{MSSA}(X_c)
$$
For the channel attention part, I follow the CBAM design:
$$
M_c = \sigma(W_1(W_0(\text{AvgPool}(X))) + W_1(W_0(\text{MaxPool}(X))))
$$
The multi-scale subspace attention partitions the channel dimension into several subspaces. Each subspace applies a 5×5 depthwise convolution followed by depthwise convolutions of varying kernel sizes (1×7 and 7×1, as well as 1×3 and 3×1). The outputs are fused via summation and combined with the original features through element-wise multiplication. A residual connection is added to ensure gradient flow:
$$
\text{MSA}(X) = \left( \sum_{k=1}^{K} \sum_{i=1}^{N} \text{ReLU}(D_k(X_i)) \right) \odot X + X
$$
Table 6 presents the effectiveness analysis of CPMSSA against other attention mechanisms.
| Model | P (%) | R (%) | Params (M) | FLOPs (G) | mAP50 (%) | mAP50-95 (%) |
|---|---|---|---|---|---|---|
| YOLO11n | 70.93 | 71.22 | 2.59 | 6.40 | 71.82 | 45.93 |
| + LRSA | 69.83 | 70.15 | 2.54 | 6.60 | 71.08 | 46.03 |
| + SEA | 72.13 | 71.29 | 2.60 | 6.40 | 72.73 | 45.29 |
| + MSCA | 71.37 | 70.43 | 2.68 | 6.70 | 72.29 | 46.11 |
| + CPMSSA | 72.58 | 71.51 | 2.49 | 6.30 | 73.31 | 47.27 |
C2f_MSCBlock Module
The second improvement targets the multi-scale modeling capability of the model. I designed a multi-scale convolution block (MSCBlock) and used it to construct the C2f_MSCBlock module. This module effectively separates feature reuse from feature transformation paths, preserving gradient flow while enhancing representational power. The MSCBlock uses factorized depthwise convolutions to reduce parameters significantly. For instance, a standard 7×7 convolution requires $7 \times 7 \times C \times C$ parameters, while the factorized version requires only $7C + 7C = 14C$ parameters.
The forward pass of C2f_MSCBlock can be formulated as:
$$
Y_0, Y_1 = \text{Split}(f_{\text{conv}}^{1 \times 1}(X))
$$
$$
X_{\text{out}} = f_{\text{conv}}^{1 \times 1}(\text{BN}(\text{Concat}(Y_0, \text{MSCBlock}(Y_1))))
$$
Table 7 compares C2f_MSCBlock with other feature extraction modules.
| Model | P (%) | R (%) | Params (M) | FLOPs (G) | mAP50 (%) | mAP50-95 (%) |
|---|---|---|---|---|---|---|
| YOLO11n | 70.93 | 71.22 | 2.59 | 6.40 | 71.82 | 45.93 |
| + C2f_DHSA | 69.83 | 70.15 | 2.89 | 7.00 | 70.35 | 44.29 |
| + C2f_SFHF | 72.13 | 71.29 | 3.11 | 6.90 | 73.82 | 46.33 |
| + C3k2_SSA | 71.37 | 70.43 | 2.81 | 7.10 | 73.69 | 45.98 |
| + C2f_MSCBlock | 72.58 | 71.51 | 2.49 | 6.30 | 74.63 | 46.37 |
AKConv Deformable Convolution
The third improvement addresses the challenge of irregular defect shapes. I introduced AKConv (Alterable Kernel Convolution), a deformable convolution module that learns dynamic sampling offsets. Unlike standard convolution with a fixed sampling grid, AKConv adapts its sampling positions based on the input content, enabling the model to better conform to the geometric structure of defects.
Given an input feature map $X$, AKConv first predicts an offset field using a auxiliary convolution:
$$
\Delta P = f_{\text{conv}}^P(X)
$$
The actual sampling coordinates are then computed as:
$$
P = P_0 + P_n + \Delta P
$$
where $P_0$ is the initial sampling grid and $P_n$ is the shape-preset bias. Since the sampling coordinates are floating-point values, bilinear interpolation is applied for feature resampling:
$$
X_{\text{offset}} = \sum_{q} g(q) \cdot X(q)
$$
Table 8 shows the comparison between AKConv and other convolution variants.
| Model | P (%) | R (%) | Params (M) | FLOPs (G) | mAP50 (%) | mAP50-95 (%) |
|---|---|---|---|---|---|---|
| YOLO11n | 70.93 | 71.22 | 2.59 | 6.40 | 71.82 | 45.93 |
| + SPConv | 69.83 | 70.15 | 2.52 | 6.40 | 73.79 | 46.47 |
| + PSConv | 72.13 | 71.29 | 2.52 | 6.40 | 73.90 | 47.23 |
| + FDConv | 71.37 | 70.43 | 2.62 | 6.20 | 74.81 | 46.77 |
| + AKConv | 72.58 | 71.51 | 2.48 | 6.30 | 74.83 | 48.51 |
Ablation Study of MCAM-Net
To systematically verify the contribution of each proposed module, I performed a comprehensive ablation study. The results are shown in Table 9. The baseline YOLO11n achieves 71.82% mAP50 and 45.93% mAP50-95. Adding each module individually brings consistent improvements. When all three modules are combined, the model achieves 77.34% mAP50 and 48.91% mAP50-95, with parameters reduced from 2.59M to 2.21M and FLOPs reduced from 6.40G to 5.70G.
| CPMSSA | MSCBlock | AKConv | Params (M) | FLOPs (G) | mAP50 (%) | mAP50-95 (%) |
|---|---|---|---|---|---|---|
| 2.59 | 6.40 | 71.82 | 45.93 | |||
| ✓ | 2.49 | 6.30 | 73.31 | 47.27 | ||
| ✓ | 2.42 | 6.00 | 74.63 | 46.37 | ||
| ✓ | 2.48 | 6.30 | 74.83 | 48.51 | ||
| ✓ | ✓ | 2.38 | 6.10 | 75.71 | 48.39 | |
| ✓ | ✓ | 2.32 | 5.90 | 76.29 | 47.31 | |
| ✓ | ✓ | 2.31 | 5.90 | 74.28 | 48.35 | |
| ✓ | ✓ | ✓ | 2.21 | 5.70 | 77.34 | 48.91 |
Comparison with State-of-the-Art Models
I benchmarked MCAM-Net against several popular detection models on the MS2025 dataset. As shown in Table 10, MCAM-Net consistently outperforms all compared methods. It achieves the highest precision of 77.85%, the highest recall of 77.20%, and the best mAP50 of 77.34% and mAP50-95 of 48.91%, all while maintaining the lowest FLOPs among the compared lightweight models.
| Model | P (%) | R (%) | Params (M) | FLOPs (G) | mAP50 (%) | mAP50-95 (%) |
|---|---|---|---|---|---|---|
| Faster R-CNN | 76.12 | 75.45 | 41.25 | 91.03 | 75.92 | 46.11 |
| FCOS | 73.28 | 72.63 | 31.85 | 78.73 | 72.82 | 44.31 |
| RetinaNet | 72.85 | 74.24 | 36.21 | 82.57 | 73.67 | 46.27 |
| RT-DETR | 73.16 | 75.52 | 19.87 | 56.90 | 74.15 | 48.33 |
| YOLOv8n | 75.95 | 74.32 | 3.01 | 8.20 | 74.29 | 46.23 |
| YOLOv9t | 75.63 | 75.98 | 2.01 | 7.90 | 73.73 | 47.38 |
| YOLOv10n | 74.03 | 76.36 | 2.71 | 8.40 | 74.18 | 46.29 |
| YOLO11n | 75.58 | 74.94 | 2.59 | 6.40 | 73.82 | 45.93 |
| YOLOv12n | 75.87 | 74.25 | 2.60 | 6.50 | 74.93 | 47.21 |
| MCAM-Net | 77.85 | 77.20 | 2.21 | 5.70 | 77.34 | 48.91 |
Generalization Study
To evaluate the generalization capability of MCAM-Net, I tested it on the EL2021 dataset used in the EL defect detection scenario. This experiment was designed to investigate whether the model could serve as a universal detector for solar panel surface defects across both imaging modalities. The results are summarized in Table 11.
| Model | P (%) | R (%) | Params (M) | FLOPs (G) | mAP50 (%) | mAP50-95 (%) |
|---|---|---|---|---|---|---|
| YOLO11n | 92.95 | 91.12 | 2.59 | 6.40 | 90.24 | 65.02 |
| EVT-Net | 92.37 | 93.27 | 6.26 | 22.39 | 92.75 | 66.73 |
| MCAM-Net | 93.16 | 93.59 | 2.21 | 5.70 | 93.76 | 67.25 |
MCAM-Net achieved the highest mAP50 of 93.76% and mAP50-95 of 67.25% on the EL2021 dataset, outperforming both YOLO11n and the EL-specific EVT-Net model. More importantly, it achieved this with substantially fewer parameters (2.21M) and FLOPs (5.70G) than EVT-Net, demonstrating that careful architectural design can yield a lightweight yet universal model for solar panel defect detection across multiple imaging modalities.
Detection System Design and Implementation
Based on the MCAM-Net model, I designed and implemented a comprehensive surface defect detection system for solar panels. The system was developed using PyQt5 as the graphical user interface framework and Python as the programming language. The system architecture follows a modular design, comprising user authentication, image and video detection, and result management functionalities.
The detection workflow begins with model weight selection. The system includes two weight configurations: one trained on EL images for production-line inspection, and one trained on visible-light images for outdoor maintenance inspection. Users can select the appropriate model based on their detection scenario. The system supports both image and video detection modes, where the selected input stream is processed through the MCAM-Net model to produce detection results. The outputs are displayed on the user interface with defect bounding boxes, class labels, and confidence scores. Users can save the detection results in CSV format for subsequent analysis.
The user interface was designed to be intuitive and user-friendly. The login interface supports both regular and administrator accounts, with registration and password modification features to ensure data security. The main interface includes model selection, image detection, video detection, start/stop detection, and result saving buttons. The system successfully implemented both EL image defect detection and visible-light image defect detection, demonstrating its effectiveness in real-world applications.
For the system validation, I conducted functional tests on both image and video detection modes. The EL image test used defect images from the EL2021 dataset, and the system successfully detected subtle crack defects with high confidence. The visible-light image test confirmed the system’s ability to detect physical damage, dust, and snow defects accurately. For video detection validation, I used drone-captured footage of a rooftop distributed photovoltaic station. The system successfully identified non-defective solar panels without producing false positives, confirming its reliability in actual inspection scenarios.
Conclusion
In this research, I have developed and validated deep learning algorithms for automated surface defect detection of polycrystalline silicon solar panels, covering both production-line and outdoor maintenance scenarios.
For the production-line EL imaging scenario, I proposed EVT-Net, which integrates three key innovations. The efficient M-CASViT backbone maintains the global modeling capability of Transformers while reducing computational cost through optimized stage configuration and the M-Stem multi-branch downsampling module. The DDC-FPN module dynamically calibrates multi-scale features, enhancing small defect detection. The decoupled detection head resolves potential conflicts between classification and regression tasks. Extensive experiments on the EL2021 dataset demonstrated that EVT-Net achieves 92.75% mAP50 and 66.73% mAP50-95, outperforming several state-of-the-art detection models.
For the outdoor visible-light imaging scenario, I proposed MCAM-Net based on YOLO11. The innovative CPMSSA attention mechanism combines channel attention with multi-scale subspace attention to enhance sensitivity to low-contrast defects. The C2f_MSCBlock module provides efficient multi-scale feature extraction with significantly reduced parameters. The AKConv deformable convolution module adapts to irregular defect shapes. The complete MCAM-Net achieved 77.34% mAP50 and 48.91% mAP50-95 on the MS2025 dataset, representing improvements of 3.52% and 2.98% over the baseline YOLO11n, while also reducing the parameter count by 0.38M. The generalization experiment further demonstrated its versatility, achieving 93.76% mAP50 on the EL2021 dataset with only 2.21M parameters, making it an excellent candidate for a universal solar panel defect detection model.
Based on MCAM-Net, I developed a complete detection system with a user-friendly interface. The system successfully supports both image and video detection modes and has been verified in real-world testing scenarios, providing effective technical support for intelligent solar panel inspection.
Future work will explore several directions. First, I plan to investigate multi-modal data fusion, combining EL, infrared thermography, and visible-light imaging to provide a more comprehensive assessment of solar panel health. Second, I will explore domain adaptation and meta-learning techniques to improve model performance under extreme weather conditions and enable online adaptation to new environments. Third, I aim to integrate temporal analysis to predict defect evolution trends and estimate the remaining useful life of solar panels, transforming defect detection into predictive maintenance for enhanced reliability and efficiency in photovoltaic power plants.