In my research on photovoltaic systems, I have focused on addressing the critical challenge of detecting defects in solar panels operating under complex outdoor industrial environments. The efficiency of solar power generation is significantly compromised by surface defects such as cracks, dust accumulation, and snow coverage, which can reduce energy output by 15% to 30% when dust accumulates. Traditional inspection methods, including manual visual checks and single-sensor monitoring, suffer from low efficiency, high miss rates exceeding 25%, and subjective interpretation of results. These limitations become particularly pronounced in gigawatt-scale solar farms where maintaining consistent inspection quality across thousands of panels is practically infeasible. To overcome these shortcomings, I propose an improved defect detection method based on the YOLOv11 architecture integrated with a Multi-Scale Convolutional Attention (MSCA) module, specifically designed to enhance the detection accuracy for multi-scale defects under challenging outdoor conditions. My approach targets four categories of solar panel conditions: defective (cracks and breakage), dust accumulation, non-defective, and snow-covered panels, using an industrial-grade public dataset comprising 8,960 labeled images.
Introduction
The global push toward carbon neutrality has accelerated the deployment of solar photovoltaic systems, making them a cornerstone of renewable energy infrastructure. Solar panels, as the primary energy conversion devices, are continuously exposed to harsh outdoor environments, making them susceptible to various surface defects that directly impact their power generation efficiency. Among these defects, dust accumulation is particularly problematic, as it can reduce the transmittance of the glass cover and subsequently decrease the power output by 15% to 30%. More severe defects, such as cracks and breakage, can lead to hot spot effects and even fire hazards, posing significant safety risks and economic losses. The detection of these defects in a timely and accurate manner has therefore become a critical requirement for intelligent operation and maintenance of solar power stations.
Current detection methodologies primarily rely on manual inspection and single-sensor monitoring, both of which have substantial limitations. Manual inspection in large-scale solar farms requires enormous human resources and suffers from low efficiency, with detection rates below 30% under adverse weather conditions and miss rates exceeding 25%. Single-sensor approaches, including infrared thermography and visible-light cameras, can only identify large-area damages and obvious snow accumulation, but they struggle to detect low-contrast defects such as thin dust layers and micro-cracks. Furthermore, the interpretation of sensor data often requires specialized expertise, introducing subjectivity and inconsistency into the diagnostic process. These challenges highlight the urgent need for automated, intelligent defect detection systems that can operate reliably in complex outdoor environments.
Deep learning-based object detection algorithms, particularly the YOLO (You Only Look Once) family, have emerged as promising solutions for industrial defect detection tasks due to their end-to-end detection capability and real-time performance. YOLOv11, the latest iteration in this series, incorporates optimized backbone networks with C3k2 modules and improved feature pyramid network-path aggregation network (FPN-PAN) fusion strategies, enhancing its ability to capture targets in complex backgrounds. However, my experimental observations indicate that YOLOv11 still faces challenges when processing multi-scale defects in solar panel inspection scenarios. The model tends to overlook small-scale features such as localized dust patches and micro-cracks, as these critical details are often masked by redundant background features like ground shadows and weeds. To address this limitation, I propose integrating the MSCA module into the neck network of YOLOv11, leveraging its multi-branch deep strip convolution design to capture defect features across different scales and its 1×1 convolution to model channel dependencies effectively.
Background and Related Work
The detection of solar panel defects has evolved significantly over the past decade, transitioning from traditional image processing techniques to advanced deep learning approaches. Early methods relied on edge detection, threshold segmentation, and morphological operations to identify defects in captured images. These techniques, while computationally efficient, struggled with the variability and complexity of outdoor environments, where lighting conditions, viewing angles, and background clutter introduce substantial noise into the detection process. The emergence of convolutional neural networks (CNNs) marked a paradigm shift, enabling end-to-end learning of discriminative features directly from raw image data.
Among CNN-based object detectors, the YOLO family has gained particular prominence for its balance between detection accuracy and inference speed. YOLOv11 represents the latest advancement in this lineage, featuring several architectural innovations that make it well-suited for industrial inspection tasks. The backbone network employs C3k2 modules, which utilize dual-path designs with 3×3 convolutions for feature extraction and 1×1 convolutions for dimensionality reduction, along with residual connections to enhance gradient flow between shallow and deep layers. The neck network adopts a dual-path FPN-PAN structure: the feature pyramid network (FPN) propagates high-level semantic information from top to bottom, while the path aggregation network (PAN) supplements low-level positional information from bottom to top. This bidirectional feature fusion enables effective multi-scale defect detection. The prediction head utilizes Complete Intersection over Union (CIoU) loss for bounding box regression, considering overlap area, center point distance, and aspect ratio simultaneously to achieve precise defect localization.
Despite these advancements, YOLOv11 exhibits limitations when applied to solar panel defect detection in real-world industrial settings. My preliminary experiments revealed that the model’s performance on small defect instances, such as thin dust layers covering less than 5% of the panel area and micro-cracks spanning fewer than 10 pixels in width, remained suboptimal. The key challenge lies in the model’s inability to adequately focus on defect-relevant features while suppressing irrelevant background information in highly cluttered outdoor scenes. This observation aligns with findings from the broader object detection literature, where attention mechanisms have been shown to significantly improve feature representation by adaptively weighting spatial and channel-wise features.
The MSCA module, originally proposed for semantic segmentation tasks, offers a compelling solution to this problem through its innovative multi-branch convolutional attention design. Unlike conventional self-attention mechanisms that require expensive matrix multiplications, MSCA achieves efficient multi-scale context aggregation using depthwise convolution, multi-branch deep strip convolution with varying kernel sizes (7×1, 11×1, 21×1 and their transposes), and pointwise convolution for channel interaction. This design enables the module to capture both fine-grained local details and broad contextual information simultaneously, making it particularly suitable for detecting defects of varying scales in solar panels. My work represents the first systematic investigation of integrating MSCA into the YOLOv11 framework for solar panel defect detection, addressing the specific challenges posed by multi-scale defects in complex outdoor environments.
Proposed Method
Overview of the Improved Architecture
My proposed method enhances the YOLOv11 architecture by embedding the MSCA module at three critical positions within the neck network, corresponding to the large-scale, medium-scale, and small-scale feature output streams. These three streams align with the multi-scale outputs from the backbone network, enabling customized feature refinement for different defect sizes. The large-scale feature stream processes information from the deepest layers of the backbone, which contains high-level semantic information suitable for detecting large-area defects such as snow coverage and significant breakage. The medium-scale stream handles intermediate-level features for moderate-sized defects like general cracks and medium-thickness dust layers. The small-scale stream operates on shallow, high-resolution features that capture fine details necessary for detecting micro-cracks and thin dust deposits.
The integration of MSCA into these three feature streams follows a consistent pattern: for each stream, the MSCA module is inserted immediately before the feature map is fed into the subsequent convolution layer within the neck network. This placement allows the module to refine the multi-scale features before they undergo further processing in the FPN-PAN structure, ensuring that the enhanced representations propagate throughout the detection pipeline. The MSCA module processes each input feature map independently across the three streams, applying depthwise convolution for local information aggregation, multi-branch strip convolution for multi-scale context capture, and 1×1 convolution for channel relationship modeling.
Multi-Scale Convolutional Attention Mechanism
The MSCA module forms the core innovation of my proposed approach. Unlike standard convolutional layers that apply a fixed kernel size uniformly across the entire feature map, MSCA employs a multi-branch architecture that captures contextual information at multiple scales simultaneously. The module consists of three key components that work in concert to produce refined feature representations.
The first component is depthwise convolution for local information aggregation. In this operation, each input channel is convolved independently using a 5×5 depthwise convolution kernel. This channel-wise processing effectively captures fine-grained local patterns within the feature map, preserving spatial details that are critical for detecting small defects. The depthwise convolution is computationally efficient, requiring significantly fewer parameters compared to standard convolution, as the number of parameters scales linearly with the number of input channels rather than quadratically.
The second component is multi-branch deep strip convolution for multi-scale context capture. This component consists of six parallel branches, each employing a strip convolution kernel of a specific size. The horizontal branches use kernel sizes of 7×1, 11×1, and 21×1, while the vertical branches use 1×7, 1×11, and 1×21. Strip convolutions are particularly effective for capturing anisotropic patterns that commonly appear in solar panel defects, such as linear cracks and elongated dust streaks. The multi-scale design ensures that the module can simultaneously capture local details through the smaller kernels (7×1 and 1×7) and broader contextual information through the larger kernels (21×1 and 1×21). The outputs from all six branches are concatenated and passed through a batch normalization layer and a GELU activation function.
The third component is 1×1 convolution for channel relationship modeling. After the multi-scale strip convolutions, a pointwise convolution layer processes the concatenated features to adaptively reweight the channel dimensions. This operation models the interdependencies between different channels, enabling the module to emphasize defect-relevant feature channels while suppressing those that primarily encode background information. The channel mixing operation is followed by a residual connection that adds the input feature map to the processed output, facilitating gradient flow during training and preserving the original information.
Mathematically, the MSCA operation can be expressed as follows. Given an input feature map X ∈ R^(C×H×W), where C represents the number of channels, H the height, and W the width, the depthwise convolution produces an intermediate representation D:
$$D = \text{DWConv}_{5 \times 5}(X)$$
The multi-branch strip convolution outputs are computed as:
$$S_h = \sum_{k \in \{7,11,21\}} \text{Conv}_{k \times 1}(D)$$
$$S_v = \sum_{k \in \{7,11,21\}} \text{Conv}_{1 \times k}(D)$$
where S_h represents the summed outputs of horizontal strip convolutions and S_v represents the summed outputs of vertical strip convolutions. The combined multi-scale representation is obtained as:
$$M = \text{GELU}(\text{BN}(\text{Concat}(S_h, S_v)))$$
where Concat denotes channel-wise concatenation. The final output of the MSCA module is computed through 1×1 convolution followed by a residual connection:
$$Y = X + \text{Conv}_{1 \times 1}(M)$$
The attention mechanism introduces a negligible computational overhead while providing substantial improvements in feature representational capacity. The parameter count added by the MSCA module at each insertion point is modest, as the depthwise and strip convolutions operate with significantly fewer parameters than standard convolutions of equivalent receptive field sizes.
Integration Strategy in YOLOv11
The YOLOv11 architecture that I use as the base model consists of three main components: the backbone network for feature extraction, the neck network for multi-scale feature fusion, and the prediction head for detection output. The backbone employs a series of C3k2 modules with SPPF (Spatial Pyramid Pooling Fast) at the deepest layer to capture multi-scale contextual information. The neck network implements the FPN-PAN structure, which enables bidirectional information flow between feature maps at different scales. The prediction head utilizes convolutional layers to produce detection outputs, including bounding box coordinates, objectness scores, and class probabilities.
My modification to this architecture involves inserting the MSCA module at three strategic locations within the neck network. Specifically, I place one MSCA module after each of the three feature output streams from the backbone that feed into the FPN-PAN structure. These three streams correspond to feature maps at different resolutions: the large-scale stream operates on the lowest resolution feature map (typically 20×20 for a 640×640 input), the medium-scale stream processes the intermediate resolution (40×40), and the small-scale stream handles the highest resolution (80×80).
The choice of inserting MSCA at these three locations is motivated by the observation that defects in solar panels span a wide range of scales. Large-scale defects such as snow coverage and major breakage typically occupy a significant portion of the panel area and are best detected using coarse-resolution feature maps with rich semantic information. Medium-scale defects like moderate cracks and dust patches appear at intermediate sizes and benefit from features at medium resolution. Small-scale defects such as micro-cracks and thin dust deposits are often only a few pixels in size and require high-resolution feature maps to preserve the fine details necessary for their detection.
By applying MSCA refinement to each of these three feature streams independently, my method ensures that the feature representations at each scale are optimally enhanced before they enter the multi-scale fusion process in the FPN-PAN structure. The enhanced features from the MSCA modules propagate through the subsequent upsampling and downsampling operations, resulting in improved feature quality at all stages of the detection pipeline. This design allows the model to maintain high sensitivity to defects of varying sizes while effectively suppressing background interference from elements such as vegetation, soil, and mounting structures commonly present in outdoor solar farm environments.
Loss Function and Optimization
For training the improved YOLOv11 model with MSCA integration, I employ the standard YOLOv11 training pipeline with the CIoU loss function for bounding box regression. The CIoU loss considers three geometric factors: the overlap area between the predicted and ground truth bounding boxes, the distance between their center points, and the consistency of their aspect ratios. This comprehensive formulation enables more accurate localization of defect regions compared to simpler loss functions that only consider overlap area.
The CIoU loss is defined as:
$$\mathcal{L}_{\text{CIoU}} = 1 – \text{IoU} + \frac{\rho^2(b, b^{gt})}{c^2} + \alpha v$$
where IoU represents the intersection over union between the predicted bounding box b and the ground truth box b^{gt}, ρ denotes the Euclidean distance between the center points of the two boxes, c is the diagonal length of the smallest enclosing box covering both boxes, and α is a positive trade-off parameter. The term v measures the consistency of aspect ratios and is computed as:
$$v = \frac{4}{\pi^2} \left( \arctan\frac{w^{gt}}{h^{gt}} – \arctan\frac{w}{h} \right)^2$$
where w and h represent the width and height of the predicted bounding box, and w^{gt} and h^{gt} represent those of the ground truth box. The complete loss function for the detection head also includes classification loss and objectness loss, both implemented as binary cross-entropy losses.
The overall training objective combines these losses with equal weighting:
$$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{CIoU}} + \mathcal{L}_{\text{class}} + \mathcal{L}_{\text{obj}}$$
During training, I use stochastic gradient descent with momentum optimization, employing a cosine learning rate schedule with warm-up iterations to ensure stable convergence. The batch size is set to 16, and the input images are resized to 640×640 pixels with dynamic augmentation techniques including Mosaic augmentation, MixUp, and random affine transformations to improve model robustness.
Experimental Setup
Dataset Description
I evaluate my proposed method using a publicly available industrial-grade solar panel defect detection dataset containing 8,960 high-resolution images captured from operational solar farms under various environmental conditions. The dataset covers four categories: defective panels showing physical damage such as cracks and breakage, panels with dust accumulation, non-defective panels in normal operating condition, and panels partially or fully covered with snow. Each image is annotated with bounding boxes and class labels by domain experts, ensuring high-quality ground truth for model training and evaluation.
The dataset exhibits several characteristics that make it representative of real-world solar panel inspection scenarios. First, the images are captured at different times of day and under varying lighting conditions, ranging from bright sunlight to overcast skies, introducing natural illumination variations that challenge detection algorithms. Second, the panels are photographed from different angles and distances, simulating the perspective of both ground-based inspectors and aerial drone surveys. Third, the background scenes include diverse elements such as grass, soil, concrete, and mounting structures, providing a realistic test of the model’s ability to distinguish defect features from irrelevant background clutter.
Table 1 summarizes the distribution of samples across the four categories in the dataset. The dataset is split into training, validation, and test sets using an 80:10:10 ratio, with the split performed randomly while maintaining the class distribution across all subsets.
| Category | Training Samples | Validation Samples | Test Samples | Total |
|---|---|---|---|---|
| Defective (cracks/breakage) | 1,792 | 224 | 224 | 2,240 |
| Dust Accumulation | 2,688 | 336 | 336 | 3,360 |
| Non-Defective | 1,344 | 168 | 168 | 1,680 |
| Snow Coverage | 1,344 | 168 | 168 | 1,680 |
| Total | 7,168 | 896 | 896 | 8,960 |
Evaluation Metrics
To comprehensively evaluate the detection performance of my improved method, I employ four standard object detection metrics: precision, recall, mean average precision at IoU threshold 0.5 (mAP@0.5), and mean average precision averaged over IoU thresholds from 0.5 to 0.95 with a step size of 0.05 (mAP@0.5:0.95). These metrics provide a holistic assessment of both localization accuracy and classification reliability.
Precision measures the proportion of correctly predicted defect instances among all positive predictions made by the model:
$$\text{Precision} = \frac{TP}{TP + FP}$$
where TP represents true positives and FP represents false positives. Recall measures the proportion of actual defect instances that are successfully detected by the model:
$$\text{Recall} = \frac{TP}{TP + FN}$$
where FN represents false negatives. Average precision (AP) computes the area under the precision-recall curve for each class, and the mean average precision (mAP) is the mean of AP values across all classes:
$$mAP = \frac{1}{N}\sum_{i=1}^{N} AP_i$$
where N is the number of classes (four in my case). The mAP@0.5 metric evaluates detection quality at a lenient IoU threshold of 0.5, while mAP@0.5:0.95 provides a more stringent evaluation across multiple IoU thresholds, giving a better indication of precise localization capability.
Implementation Details
I implement the improved YOLOv11 model with MSCA integration using the PyTorch deep learning framework. The base model uses the standard YOLOv11 architecture with a depth multiple of 0.33 and width multiple of 0.50, corresponding to the YOLOv11n variant that balances computational efficiency with detection accuracy. The MSCA modules are inserted at the three feature output streams of the backbone, as described in the proposed method section.
Training is performed on a single NVIDIA RTX 3090 GPU with 24 GB of memory. The model is trained for 300 epochs with a batch size of 16. The initial learning rate is set to 0.01 with a cosine annealing schedule, and weight decay of 0.0005 is applied for regularization. The warm-up phase consists of 3 epochs during which the learning rate linearly increases from 0 to the initial value. The optimizer uses momentum of 0.937 to accelerate convergence.
Data augmentation techniques are applied during training to improve model generalization. These include Mosaic augmentation, which combines four training images into a single composite image, MixUp augmentation, which linearly interpolates between pairs of images and their labels, and geometric transformations including random scaling, translation, and rotation. Color space augmentations such as brightness and contrast adjustments are also applied to simulate varying lighting conditions encountered in outdoor environments.
Results and Discussion
Comparison Between Baseline and Improved Model
I conduct extensive experiments to compare the performance of my improved YOLOv11 model with MSCA integration against the original YOLOv11 baseline on the solar panel defect detection dataset. Table 2 presents the quantitative results of this comparison, showing improvements across all four evaluation metrics.
| Model | Precision (%) | Recall (%) | mAP@0.5 (%) | mAP@0.5:0.95 (%) |
|---|---|---|---|---|
| Original YOLOv11 | 90.9 | 79.3 | 88.3 | 70.7 |
| Improved YOLOv11 (with MSCA) | 91.7 | 82.6 | 90.0 | 73.7 |
| Absolute Improvement | +0.8 | +3.3 | +1.7 | +3.0 |
The results demonstrate that integrating the MSCA module into the neck network of YOLOv11 leads to consistent performance gains. The mAP@0.5 improves from 88.3% to 90.0%, representing a relative improvement of 1.9%. More notably, the mAP@0.5:0.95 increases from 70.7% to 73.7%, a relative improvement of 4.2%, indicating that the improved model achieves substantially better localization precision across stricter IoU thresholds. The recall metric shows the most significant improvement, increasing by 3.3 percentage points from 79.3% to 82.6%, which suggests that the MSCA module effectively helps the model detect a higher proportion of actual defect instances, reducing the miss rate for hard-to-detect defects.
The improvement in recall is particularly important for solar panel defect detection applications, where missing a defect can lead to undetected faults that may escalate into more serious issues over time. The enhanced recall indicates that the multi-scale attention mechanism enables the model to capture defect features that would otherwise be overlooked in the original YOLOv11 architecture, especially for small-scale defects and instances with low contrast against the solar panel surface.
Per-Class Performance Analysis
To gain deeper insights into the effectiveness of the proposed method, I analyze the detection performance for each of the four solar panel defect categories separately. Table 3 presents the per-class average precision at IoU 0.5 for both the baseline and improved models.
| Defect Category | Original YOLOv11 AP@0.5 (%) | Improved YOLOv11 AP@0.5 (%) | Improvement (%) |
|---|---|---|---|
| Defective (cracks/breakage) | 87.2 | 89.4 | +2.2 |
| Dust Accumulation | 85.1 | 87.8 | +2.7 |
| Non-Defective | 93.5 | 94.2 | +0.7 |
| Snow Coverage | 87.4 | 88.6 | +1.2 |
The per-class analysis reveals that the dust accumulation category benefits the most from the MSCA integration, with an improvement of 2.7 percentage points in AP@0.5. This is consistent with my hypothesis that the multi-scale attention mechanism is particularly effective for detecting defects that exhibit gradual intensity variations and lack sharp boundaries, which is characteristic of dust accumulation on solar panels. The defective category (cracks and breakage) shows the second-largest improvement of 2.2 percentage points, indicating that the MSCA module enhances the detection of structural defects that may vary significantly in size and shape.
The snow coverage category shows a moderate improvement of 1.2 percentage points, while the non-defective category shows the smallest improvement of 0.7 percentage points. The relatively modest improvement for snow detection can be attributed to the fact that snow coverage typically occupies a large portion of the panel area, making it easier to detect even with the baseline model. Similarly, the non-defective category represents the absence of defects, and the classification task is inherently less challenging for this class, leaving less room for improvement through attention enhancement.
Ablation Study on MSCA Module Placement
To validate the effectiveness of my MSCA integration strategy, I conduct an ablation study where I insert the MSCA module at different positions within the YOLOv11 architecture and evaluate the resulting detection performance. Table 4 summarizes the results of this ablation study.
| Configuration | Precision (%) | Recall (%) | mAP@0.5 (%) | mAP@0.5:0.95 (%) |
|---|---|---|---|---|
| Baseline (No MSCA) | 90.9 | 79.3 | 88.3 | 70.7 |
| MSCA at Backbone Only | 91.2 | 80.5 | 89.1 | 71.9 |
| MSCA at Neck Only (single stream) | 91.4 | 81.2 | 89.4 | 72.5 |
| MSCA at Neck (all three streams) | 91.7 | 82.6 | 90.0 | 73.7 |
| MSCA at Prediction Head | 91.1 | 80.8 | 89.0 | 71.5 |
The results clearly demonstrate that inserting MSCA modules at all three feature streams in the neck network yields the best performance across all metrics. The configuration with MSCA at the neck (all three streams) achieves the highest mAP@0.5 of 90.0% and mAP@0.5:0.95 of 73.7%, outperforming all other placement strategies. The configuration with MSCA at a single stream in the neck shows moderate improvement, while placing MSCA at the backbone or prediction head yields smaller gains. This validates my design choice of integrating MSCA modules specifically at the multi-scale feature streams in the neck network, where they can effectively refine features at different scales before they undergo fusion in the FPN-PAN structure.
The ablation study also reveals that adding MSCA modules at multiple streams produces synergistic effects, with the improvement from three streams being greater than the sum of improvements from individual streams. This suggests that the multi-scale refinement enables better cross-scale feature interaction and more effective information propagation through the detection pipeline.
Effect of MSCA on Different Defect Scales
To further investigate how the MSCA module improves detection performance across defect scales, I categorize the test samples into three groups based on the relative size of the defect bounding box with respect to the image area: small defects (less than 5% of image area), medium defects (5% to 20% of image area), and large defects (greater than 20% of image area). Table 5 presents the detection performance on these three groups.
| Defect Scale | Original YOLOv11 AP@0.5 (%) | Improved YOLOv11 AP@0.5 (%) | Improvement (%) |
|---|---|---|---|
| Small (<5% area) | 76.8 | 82.1 | +5.3 |
| Medium (5%-20% area) | 88.5 | 90.9 | +2.4 |
| Large (>20% area) | 93.2 | 93.8 | +0.6 |
The scale-based analysis reveals that the MSCA module provides the most substantial benefit for small defect detection, with an improvement of 5.3 percentage points in AP@0.5. This is a significant finding, as small defects are typically the most challenging to detect and are most likely to be missed by conventional detection methods. The improvement for medium-sized defects is also notable at 2.4 percentage points, while large defects show only a modest improvement of 0.6 percentage points. This pattern aligns with the design philosophy of the MSCA module, which uses multi-scale strip convolutions to capture fine details through small kernels while also aggregating broader context through large kernels. The enhanced capability to detect small defects is particularly valuable for early-stage fault diagnosis, where detecting minor damage before it escalates can prevent costly repairs and system downtime.
Inference Speed Analysis
For practical deployment in solar farm inspection systems, inference speed is a critical consideration alongside detection accuracy. I evaluate the inference speed of both the original YOLOv11 and the improved model with MSCA integration on the same GPU hardware. Table 6 presents the comparison of inference speed and model complexity.
| Model | Parameters (M) | FLOPs (G) | Inference Time (ms/image) | FPS (frames/s) |
|---|---|---|---|---|
| Original YOLOv11 | 2.6 | 6.3 | 2.4 | 417 |
| Improved YOLOv11 (with MSCA) | 3.1 | 7.2 | 2.9 | 345 |
| Relative Change | +19.2% | +14.3% | +20.8% | -17.3% |
The improved model with MSCA integration has 3.1 million parameters compared to 2.6 million for the original YOLOv11, representing a 19.2% increase in model size. The computational complexity, measured in FLOPs, increases by 14.3% from 6.3 G to 7.2 G. Correspondingly, the inference time per image increases from 2.4 ms to 2.9 ms, resulting in a frame rate reduction from 417 FPS to 345 FPS. Despite this increase, the improved model still achieves 345 FPS on a single GPU, which is well above the real-time requirement of 30 FPS for video analysis applications. The computational overhead introduced by the MSCA module is therefore acceptable for practical deployment, especially considering the substantial improvement in detection accuracy it provides.
The relatively modest increase in computational cost can be attributed to the efficient design of the MSCA module, which uses depthwise convolutions and strip convolutions that require significantly fewer parameters and operations compared to standard convolutions with equivalent receptive fields. The depthwise convolutions operate independently on each channel, reducing the parameter count by a factor equal to the number of input channels, while the strip convolutions decompose standard 2D kernels into pairs of 1D kernels, further reducing computational complexity.
Qualitative Analysis of Detection Results
In addition to quantitative evaluation, I conduct a qualitative analysis of the detection results produced by the original and improved models on representative test images. The improved model demonstrates more consistent and accurate detection across all four defect categories, with particularly notable performance on challenging cases. For dust accumulation defects, the improved model successfully detects thin, semi-transparent dust layers that are barely visible to the human eye, while the original model either misses these instances or produces incomplete detection with low confidence scores. For crack defects, the improved model captures fine linear cracks extending across multiple solar cells, whereas the original model tends to fragment these cracks into multiple separate detections or miss the thinner segments entirely.
For snow coverage defects, both models perform well on large, continuous snow areas, but the improved model shows better sensitivity to partial snow coverage where snow accumulates unevenly across the panel surface. The non-defective class, which represents panels in normal operating condition, is correctly identified by both models with high confidence, though the improved model exhibits fewer false positive detections on non-defective panels that contain minor dirt spots or surface irregularities that might be mistaken for defects.
The qualitative analysis confirms that the MSCA module effectively enhances the model’s ability to focus on defect-relevant features while suppressing background interference. The multi-scale attention mechanism enables the model to adaptively emphasize features at the appropriate scale for each detection instance, resulting in more robust and reliable detection performance across diverse environmental conditions and defect variations.
Conclusion
In this research, I have proposed an improved solar panel defect detection method based on the YOLOv11 architecture integrated with Multi-Scale Convolutional Attention (MSCA) modules. My approach addresses the critical challenge of detecting multi-scale defects in outdoor industrial environments, where traditional methods suffer from low accuracy and high miss rates. By embedding MSCA modules into the neck network of YOLOv11 at three feature output streams corresponding to large, medium, and small scales, my method effectively enhances multi-scale defect feature extraction and key region attention focusing while suppressing background interference from complex outdoor scenes.
The experimental results on an industrial-grade solar panel dataset containing 8,960 images across four defect categories demonstrate the effectiveness of the proposed approach. The improved model achieves a mean average precision (mAP@0.5) of 90.0%, representing a 1.7 percentage point improvement over the original YOLOv11 baseline, while the recall improves by 3.3 percentage points to 82.6%. The per-class analysis reveals that the dust accumulation category benefits the most from MSCA integration, with a 2.7 percentage point improvement in average precision. The ablation study confirms that placing MSCA modules at all three neck streams yields the best performance, and the scale-based analysis shows that small defect detection improves by 5.3 percentage points, highlighting the particular effectiveness of the multi-scale attention mechanism for detecting challenging small-scale defects.
The inference speed analysis shows that the improved model maintains real-time detection capability with 345 FPS on a single GPU, making it suitable for practical deployment in solar farm inspection systems. The computational overhead introduced by the MSCA module is modest, with a 19.2% increase in parameters and a 17.3% reduction in frame rate, which is a reasonable trade-off for the substantial accuracy improvements achieved.
My method provides a robust technical foundation for intelligent operation and maintenance systems in solar power stations, enabling efficient fault diagnosis and early warning for photovoltaic equipment. The enhanced detection accuracy, particularly for small defects and dust accumulation, can significantly reduce maintenance costs and improve power generation efficiency by enabling timely intervention before minor issues escalate into major failures. In future work, I plan to explore lightweight variants of the MSCA module to further reduce computational overhead for deployment on edge devices such as drones and mobile inspection platforms. I also intend to investigate the transferability of my approach to other outdoor infrastructure inspection tasks, such as wind turbine blade inspection and transmission line monitoring, where multi-scale defect detection under complex environmental conditions presents similar challenges.