In the context of global energy shortages and environmental concerns, solar power generation has garnered significant attention as a renewable and clean energy source. Solar panels are critical components of photovoltaic power stations, and their operational condition directly impacts the efficiency and stability of the entire system. Therefore, rapid and effective monitoring of solar panel defects is of paramount importance. Traditional inspection methods, such as electroluminescence detection, laser scanning, and manual visual inspection, are often time-consuming, labor-intensive, and prone to human error. With technological advancements, unmanned aerial vehicles (UAVs) equipped with thermal cameras have become popular for large-scale solar farm inspections, offering high efficiency and reduced risk. However, accurately identifying and extracting defect areas, such as hot spots, from thermal infrared images remains challenging due to issues like low contrast, noise, and non-uniform backgrounds.
This article proposes an improved image processing method based on the HSV color space model and an enhanced K-means clustering algorithm for detecting defects in solar panels. The method aims to precisely locate and quantify hot spot areas in thermal infrared images, enabling efficient assessment of solar panel health. The approach involves several key steps: converting infrared images to HSV space, applying bilateral filtering for noise reduction and contrast enhancement, estimating the grayscale probability density function using a Gaussian kernel, determining initial cluster centers based on prior knowledge, and performing K-means clustering to segment and extract defect regions. The proposed method demonstrates high accuracy, sensitivity, and stability in detecting solar panel defects, making it suitable for practical applications in solar farm maintenance.
Thermal infrared images capture the temperature distribution of solar panels, where defective areas, such as hot spots, exhibit higher temperatures and correspondingly different grayscale values compared to normal regions. However, these images often suffer from low contrast, noise, and irrelevant thermal signatures from non-active components like frames and busbars. To address these challenges, the proposed method begins with image preprocessing in the HSV color space. Unlike the RGB model, the HSV space separates color information (Hue and Saturation) from brightness (Value), making it more intuitive and less affected by lighting variations. The conversion from RGB to HSV is defined as follows:
$$ H = \begin{cases} 60^\circ \times \left( \frac{G – B}{\Delta} \right) & \text{if } Max = R \\ 60^\circ \times \left( 2 + \frac{B – R}{\Delta} \right) & \text{if } Max = G \\ 60^\circ \times \left( 4 + \frac{R – G}{\Delta} \right) & \text{if } Max = B \end{cases} $$
$$ S = \begin{cases} 0 & \text{if } Max = 0 \\ \frac{\Delta}{Max} & \text{otherwise} \end{cases} $$
$$ V = Max $$
where \( Max \) is the maximum value among the RGB components, and \( \Delta = Max – Min \). After conversion, bilateral filtering is applied to the Saturation (S) component to reduce noise while preserving edges. Bilateral filtering combines spatial proximity and pixel intensity similarity, effectively enhancing contrast and suppressing noise without blurring critical details.
The core of the defect detection lies in clustering the preprocessed image based on grayscale characteristics. Traditional K-means clustering is widely used but suffers from limitations such as random initialization of cluster centers, which can lead to unstable results and local optima. To overcome this, the proposed method incorporates prior knowledge derived from the grayscale probability density function of the infrared image. First, the grayscale probability density \( f(x) \) is estimated from the image data. For an image with \( N \) pixels, the probability density at grayscale value \( x \) is given by:
$$ f(x) = \frac{N_x}{N_{\text{sum}}}, \quad x \in [0, 255] $$
where \( N_x \) is the number of pixels with grayscale \( x \), and \( N_{\text{sum}} \) is the total number of pixels. To obtain a smooth density function suitable for analysis, a non-parametric estimation using the Parzen window with a Gaussian kernel is employed. The estimated probability density \( F(x) \) is:
$$ F(x) = \frac{1}{Nh} \sum_{i=1}^{N} K \left( \frac{x – x_i}{h} \right) $$
with the Gaussian kernel defined as:
$$ K(u) = \frac{1}{\sqrt{2\pi}} \exp\left(-\frac{u^2}{2}\right) $$
Here, \( x_i \) represents the grayscale values of the pixels, \( h \) is the bandwidth parameter, and \( N \) is the total number of pixels. This estimation smooths the density curve, highlighting key features such as peaks and valleys that correspond to distinct regions in the solar panel image.
The number of clusters \( K \) in K-means clustering is determined using the elbow method, which evaluates the sum of squared errors (SSE) for different \( K \) values. The SSE is calculated as:
$$ \text{SSE} = \sum_{i=1}^{K} \sum_{q \in C_i} \| q – m_i \|^2 $$
where \( C_i \) is the set of pixels in cluster \( i \), \( q \) is a pixel in \( C_i \), and \( m_i \) is the centroid of cluster \( i \). The optimal \( K \) is chosen at the “elbow” point where the rate of decrease in SSE slows significantly. Once \( K \) is determined, initial cluster centers are selected based on the estimated grayscale probability density. The density function’s extreme points (peaks and troughs) are identified, and the grayscale range \( [\theta_{\text{min}}, \theta_{\text{max}}] \) is divided into \( K \) intervals. Within each interval, the grayscale value with the highest probability density is chosen as an initial cluster center, ensuring that the centers are representative of the image’s inherent structure.
With initial centers set, the K-means algorithm proceeds iteratively. Each pixel is assigned to the cluster whose centroid is closest in Euclidean distance:
$$ \text{argmin}_{C_i} \, E_{\text{dist}}(C_i, \theta) = \| \theta – \theta_{C_i} \|^2 $$
where \( \theta \) is the pixel’s grayscale value, and \( \theta_{C_i} \) is the centroid of cluster \( C_i \). After assignment, centroids are updated by computing the mean grayscale value of all pixels in each cluster:
$$ \theta_{C_i}’ = \frac{1}{N_i} \sum_{x \in C_i} x $$
where \( N_i \) is the number of pixels in cluster \( C_i \). The process repeats until centroids converge, minimizing the SSE. This improved initialization leads to faster convergence and more stable clustering results compared to random initialization.
Post-clustering, morphological operations are applied to refine the defect regions. Since solar panels consist of multiple components (e.g., cells, busbars, frames) that may exhibit temperature variations, small irrelevant regions are removed based on area thresholds. The area of non-generating parts is estimated, and regions smaller than a threshold \( T \) are eliminated, ensuring that only genuine hot spot defects are retained. This step enhances the accuracy of defect quantification.
To evaluate the proposed method, experiments were conducted using thermal infrared images of solar panels captured by a UAV-mounted camera. The images contained various defect types, including hot spots, with challenges such as noise and low contrast. The performance was compared against traditional K-means clustering and other existing methods, such as Otsu thresholding and B-spline least squares fitting. Key metrics included accuracy, precision, recall, and F-measure, defined as:
$$ \text{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN} $$
$$ \text{Precision} = \frac{TP}{TP + FP} $$
$$ \text{Recall} = \frac{TP}{TP + FN} $$
$$ \text{F-measure} = \frac{2 \cdot TP}{2 \cdot TP + FP + FN} $$
where \( TP \) is true positives (correctly identified defects), \( TN \) is true negatives, \( FP \) is false positives, and \( FN \) is false negatives.
| Algorithm | Accuracy (%) | Precision (%) | Recall (%) | F-measure (%) |
|---|---|---|---|---|
| Proposed Improved K-means | 90.86 | 95.95 | 85.54 | 90.45 |
| Traditional K-means | 84.92 | 82.67 | 81.58 | 82.12 |
| B-spline Least Squares Fitting | 85.79 | 88.41 | 78.21 | 82.99 |
The results demonstrate that the proposed method achieves superior performance across all metrics, with an accuracy of 90.86% and an F-measure of 90.45%. Notably, it maintains high precision (95.95%) while ensuring reasonable recall (85.54%), indicating effective defect detection with minimal false positives. In contrast, traditional K-means and other methods show lower accuracy and stability due to sensitivity to noise and initialization issues.
Additionally, the computational efficiency of the proposed method was assessed by comparing iteration counts. Traditional K-means, with random initialization, exhibited variable iteration numbers ranging widely, leading to unpredictable computation times. The improved method, with data-driven initialization, consistently converged in fewer iterations (e.g., 18 iterations in tests), enhancing speed and reliability. This is crucial for large-scale solar farm inspections where thousands of solar panel images need rapid processing.
| Algorithm | Average Iterations | Iteration Range | Stability |
|---|---|---|---|
| Proposed Improved K-means | 18 | Fixed | High |
| Traditional K-means | Variable (e.g., 22-48) | 26 | Low |
The proposed method’s robustness stems from its integration of HSV space preprocessing and probability density-based initialization. By operating on the Saturation component, the method emphasizes contrast differences between defective and normal areas of the solar panel. The Gaussian kernel estimation smooths the grayscale distribution, mitigating the impact of noise and discrete points. Furthermore, the elbow method ensures an optimal number of clusters, adapting to varying image characteristics. These features collectively enable accurate segmentation of hot spots, even in challenging conditions.
In practical applications, this method can be integrated into automated inspection systems using UAVs. After capturing thermal images, the algorithm processes each solar panel image, identifies defect regions, and quantifies the damage extent, such as the percentage of affected area. This information aids maintenance teams in prioritizing repairs, reducing downtime, and optimizing energy output. The method’s adaptability makes it suitable for different types of solar panels, including monocrystalline, polycrystalline, and thin-film variants, as long as thermal signatures are distinct.
Future work could explore real-time implementation on embedded systems for onboard UAV processing, reducing data transmission needs. Additionally, combining thermal data with visible-light images or electroluminescence data might enhance defect classification, distinguishing between hot spots, cracks, and soiling. Machine learning techniques could be incorporated to further improve detection accuracy for subtle defects. However, the current method offers a solid foundation for reliable solar panel inspection without requiring extensive training data.
In conclusion, the improved K-means clustering method based on HSV space modeling provides an effective solution for detecting defects in solar panels. It addresses limitations of traditional approaches by incorporating prior knowledge from grayscale probability densities, leading to stable and accurate clustering. The method excels in handling noisy, low-contrast infrared images, accurately locating hot spots, and quantifying damage. With high performance metrics and computational efficiency, it is a valuable tool for maintaining the health and efficiency of solar power installations, contributing to the sustainable generation of clean energy.