PV Panels Fault Detection Video Method Based on Mini-Patterns

Abstract

The development of solar technologies and the widespread adoption of photovoltaic (PV) panels have significantly transformed the global energy landscape. PV panels have evolved from niche applications to become a primary source of electricity generation, driven by their environmental benefits and declining costs. However, the performance and operational lifespan of PV systems are often compromised by various faults, which can lead to efficiency losses and increased maintenance costs. Consequently, effective and timely fault detection methods have become a critical focus of current research in the field. This work proposes an innovative video-based method for the dimensional evaluation and detection of malfunctions in solar panels, utilizing processing techniques applied to aerial images captured by unmanned aerial vehicles (drones). The method is based on a novel mini-pattern matching algorithm designed to identify specific defect features despite challenging environmental conditions such as strong gradients of non-uniform lighting, partial shading effects, or the presence of accidental deposits that obscure panel surfaces. The proposed approach aims to enhance the accuracy and reliability of fault detection, enabling more efficient monitoring and maintenance of PV installations.

1. Introduction

The widespread adoption of photovoltaic (PV) technology over the past three decades marks not only a transformative technological trend but also a substantial opportunity for advancements in applied mathematics and data-driven analysis. As solar energy plays an increasingly role in both global and regional energy systems, mathematical modeling becomes essential for optimizing system design, forecasting energy output, and assessing fault risks and economic viability.

This rapid growth in PV deployment has been fueled by a combination of policy support, economic incentives, and continuous technological innovation. Developments in PV cell architectures, such as thin-film, organic, perovskite-based, and tandem silicon technologies, have significantly enhanced power conversion efficiencies while reducing production costs [1]. This progress is the culmination of nearly two centuries of research and innovation. Understanding this technological evolution—from the early discovery of the photovoltaic effect to the recent advances in multi-junction solar cells—provides essential context for analyzing current deployment trends and for developing robust models to optimize PV utilization.

The photovoltaic effect, initially observed by Alexandre Edmond Becquerel in 1839 [2], underpins all solar cell technologies. Early PV devices used materials such as selenium and tellurium. In 1883, Charles Fritts created one of the first selenium solar cells, demonstrating light-dependent resistance and achieving 1–2% efficiency [3]. Edward Weston later patented a device for converting solar to electrical and mechanical energy [4]. A major milestone came in 1940 when Russell Ohl developed the silicon p-n junction solar cell, achieving ~1% efficiency and establishing silicon’s dominance in PV technology [5]. By the 1950s, efficiency reached 6%, and Hoffman Electronics improved it to 15% in 1961 [6]. However, single-junction cells face theoretical efficiency limits due to spectral mismatch and thermalization losses. Given the bandgap energy of crystalline silicon (approximately 1.12 eV), only photons with wavelengths shorter than 1.1 μm are absorbed effectively to generate electron–hole pairs. Photons with longer wavelengths lack sufficient energy for band-to-band transitions and therefore are not absorbed, reducing the overall conversion efficiency [2]. To address this, multi-junction solar cells were developed, incorporating multiple semiconductor layers to capture broader spectral ranges. Triple-junction cells became widely used, and by the 2000s, five-junction cells reached 35.8% efficiency in space and 38.8% terrestrially [7]. As of June 2025, the highest recorded efficiency for a solar cell remains 47.1%, achieved by a six-junction inverted metamorphic solar cell under concentrated light conditions. This record was established in 2020 by researchers at the U.S. National Renewable Energy Laboratory (NREL) and remains the benchmark for multi-junction solar cell performance [8].

As PV systems evolve toward higher efficiencies and more intricate architectures, the sensitivity to manufacturing defects, material degradation, and environmental stressors increases accordingly. This complexity necessitates advanced fault detection and diagnostic methods capable of identifying subtle and diverse malfunction patterns that can significantly impact the overall energy yield and system longevity.

Advancements in hot spot detection for photovoltaic (PV) systems have shifted from conventional electrical parameter monitoring to video-based detection methods utilizing infrared (IR) image analysis [9]. Traditional approaches that rely on electrical parameters are often impractical for large-scale solar farms due to the complexity and resource demands of auxiliary circuitry. In contrast, video-based methods offer higher detection efficiency and do not require additional physical devices.

Infrared (IR) imaging enables the monitoring of temperature distribution across solar cells under varying operating conditions. By applying specialized video processing functions, hot spots (localized areas of excessive heating) can be effectively identified. Two main categories of IR image processing techniques are employed: classical image processing methods and neural network-based approaches. Among the classical techniques, binary thresholding is the most commonly used, enabling image segmentation based on a predefined threshold value [10,11]. Other methods include k-means clustering for pixel intensity grouping [12], fuzzy rule-based classification [13], color image descriptors [14], and iterative segmentation based on local intensity maxima [15]. Pattern matching techniques are also used, wherein known hot spot patterns, stored in a reference database, are searched using sliding window methods [16].

More recent approaches employ deep learning techniques, particularly convolutional neural networks (CNNs). One widely adopted model is the VGG-19 architecture [17], which performs hot spot segmentation through transfer learning and utilizes small convolution filters (3 × 3). Additional models include Faster R-CNN [18] and the RetinaNet model based on ResNet-50 and ResNet-152 residual CNN [19].

Despite the high detection accuracy achieved through neural networks, challenges remain due to discrepancies between training and testing datasets [20]. Consequently, hot spot detection continues to represent an open and evolving research problem. Therefore, the development of sophisticated image processing techniques, such as the mini-pattern matching algorithms applied to drone-captured data, becomes critical. These methods enable the accurate, large-scale monitoring of PV installations, ensuring that the high-efficiency gains achieved in cutting-edge solar cells translate effectively into real-world performance and reliability.

2. Methods

In practical applications, existing methods for evaluating hot spots in photovoltaic panels primarily focus on physically identifying their presence or estimating their dimensions in pixel units through video analysis. This paper introduces an innovative approach that enables the measurement of hot spot dimensions in millimeters, based on the numerical processing of images taken via a drone. The proposed analysis technique, grounded in advanced pattern recognition, enables partial pattern matching and robust reconstruction, even in the presence of noise or occlusion.

The mini-pattern algorithm extends traditional ones by dividing a reference pattern into multiple sub-patterns (mini-patterns) and independently evaluating their match scores within the analyzed image sequence. Unlike standard techniques that assess the entire pattern at once, this approach improves robustness by enabling partial matches. If any mini-pattern aligns correctly, the original pattern can be reconstructed based on the known relative positions of the sub-patterns.

In Figure 1, under the standard pattern matching routine, if the window (F) from the investigated image (I)—the region in which the pattern (P) is expected to match—is affected by perturbing factors, the matching score will not validate the position of that window, even though certain parts of the pattern might still align with the image. However, if the evaluation is performed only on the unaffected portions, the window position could be correctly validated as a match. This is the fundamental principle of the proposed mini-pattern routine [21].

Segmenting the initial pattern into a set of smaller sub-patterns allows for the independent evaluation of each component. If even one mini-pattern yields a valid matching score, the original pattern can be reconstructed. Reconstruction is made possible because, during the segmentation stage, the relative positioning of each mini-pattern within the original pattern is preserved.

Once the correlation between the temperature histogram and the surface area of hot spots is established, it becomes possible to precisely identify

• The number of defective points;
• Their individual and cumulative surface area;
• Their spatial distribution across the panel surface;
• Their possible causes through technical correlations.

This segmentation increases resistance to disturbances or obstruction. While conventional methods may fail if parts of the pattern are corrupted, the mini-pattern strategy validates matching based on uncorrupted components. Consequently, it replaces a single pattern matching function with multiple interrelated mini-pattern evaluations, enhancing detection reliability even under non-ideal imaging conditions.

Since image acquisition is conducted from varying altitudes relative to the PV panel surface, accurate dimensional measurements require a reference of known size on the panel itself. This dimensional reference, introduced as input data, is the panel’s physical length, as specified in its technical documentation. To locate this reference within the acquired imagery, the mini-pattern detection technique is applied.

The fault detection method is organized into two primary components (as represented in Figure 2:

• Visible spectrum processing. I(x,y) is the image captured by a drone-mounted camera. The reference panel is represented by a set of n mini-patterns ${\left\{P_i\right\}}_{i=1}^n$ extracted from a reference panel template. The image processing sequence proceeds as follows:

  1.1.

  Feature matching: Each mini-pattern Pi is searched in the input image I using the similarity index (cross-correlation). Matching scores $S_i\in \left[\mathrm{0,1}\right]$ are computed as:

  S_i = sim (P_i, I)

  (1)

  and a match is validated when S_i > τ, where τ is a predefined threshold. For this study, a threshold of 0.8 was empirically selected for the mini-pattern matching step, based on validation experiments across a range of outdoor drone-acquired images. This value was found to offer a reliable balance between matching robustness and false-positive suppression under typical disturbances such as uneven lighting, partial shading, or panel surface contamination.

  1.2.

  Panel boundary reconstruction: If at least k valid mini-patterns are identified (k < n), the algorithm reconstructs the panel geometry by estimating the transformation T (via homography) that maps the panel model coordinates (x_p, y_p) onto the image coordinates (x_i, y_i):

  $$T:(x_p,y_p)\to (x_i,y_i)$$

  (2)

  This transformation is computed using last-squares fitting.

  1.3.

  Metric calibration: If L_dim is the known length of the panel (usually 1650 mm) and the panel’s length in pixels L_pix is determined from the transformation T, the image scale factor α is calculated as:

  $$\propto ={\displaystyle \frac{L_{dim}}{L_{pix}}}\left({\displaystyle \frac{mm}{pixels}}\right)$$

  (3)

  This scale factor will be used to convert thermal feature sizes from pixels to millimeters.

2.

Infrared spectrum processing. Let I_IR(x,y) denote the infrared image aligned with I. The panel area, as determined from visible spectrum processing, is mapped onto the IR domain using the transformation T. Within this region, the following are computed:

2.1.

Thermal segmentation: A thermal gradient map is constructed as:

$$G\left(x,y\right)={\displaystyle \frac{d{I}_{IR}}{dT}}$$

(4)

This spatial localization ensures that hot spot analysis is confined strictly to the actual surface of the panel, eliminating interference from surrounding objects or external thermal anomalies. The reference temperature used in this process corresponds to a region presumed to be fault-free, typically one of the panel’s edges, which serves as a baseline for evaluating thermal deviations. Using the resulting temperature histogram, hot spot regions are isolated. As an example, Figure 3 presents comparative histograms for a fault-free panel where the main temperature area is indicated by a blue callout, in contrast to four representative panels affected by prevalent defect categories, with hot spot regions indicated by an orange callout. It can be observed that the main temperature cluster (on healthy cells) is compact and central, and the defects produce hotter peaks, arising as an individual entity on the right side of the histogram. Thus, in the case of partial shading, which causes a cooler temperature, they will produce distinct and separate histogram peaks, allowing defect identification despite the shading.

2.2.

Dimensional measurement: For each segmented hot spot, its pixel dimensions (width w_pix and height h_pix) are extracted. Using the scale factor α, obtained during visible-spectrum calibration, the surface area of the detected hot spots is determined in metric units as:

A = (w_pix × h_pix) × α²

(5)

Then, the damage rate R_d% of the panel, whose known physical surface area is A_dim, can be calculated as the ratio:

$$R_{d\%}={\displaystyle \frac{{\sum }_{i=1}^m{A}_m}{A_{dim}}}\times 100$$

(6)

By segmenting the reference pattern into smaller mini-patterns, this method allows for independent matching and increased robustness against occlusion or partial visibility. Even when only a single mini-pattern is successfully detected, the original panel pattern can be reconstructed, thus ensuring system resilience and the continuity of analysis under suboptimal conditions.

Accurate pixel-to-millimeter conversion requires that the image of the photovoltaic panel be captured along an axis normal to the panel surface—that is, perpendicular to the intersection of its diagonals. Thus, the use of a known dimensional reference enables precise calibration between pixel and real-world measurements. This reference consists of two adjacent panel edges arranged at a right angle (the panel’s known length and width), as specified in the construction data. In cases where the camera deviates from the normal axis, geometric distortions may occur. To address this, a correction routine is implemented, using the same orthogonal reference dimensions to compensate for positional deviations and restore measurement accuracy.

An important component of the proposed methodology involves the linear measurement of detected hot spots. Using the known physical dimensions of the photovoltaic panel (length and width) as input data, a correction algorithm is implemented. This algorithm computes a correction factor by comparing the aspect ratio (length-to-width) observed in the captured image, expressed in pixels, with the actual dimensional ratio from the panel’s specifications. The resulting factor is then used to adjust pixel-to-millimeter scaling along the X and Y axes for accurate defect size estimation.

The methodology also incorporates a temperature histogram calibration procedure to ensure accurate correlation between thermal data and hot spot surface areas. This process accounts for various external influences such as illumination gradients, partial shading, mechanical impact, or surface contamination. A reference region of the panel—assumed to be defect-free and typically located along one of the panel’s edges—is used to establish a baseline temperature. By comparing this reference under nominal and current conditions, the system can accurately identify thermal anomalies and determine their distribution, surface coverage, and physical dimensions. To mitigate the risk of the reference region itself containing a fault, the method includes a preliminary consistency check that evaluates local thermal uniformity and excludes outliers. Additionally, the algorithm leverages the repetitive structure of PV modules to cross-validate patterns across multiple cells, improving robustness even if minor deviations occur within the chosen reference area. Thus, the risk of misclassification due to a compromised baseline is reduced, and the detection remains sensitive to actual faults.

3. Results

In order to test the proposed method, 12 photovoltaic panels with four types of defects (junction box, single-cell, multicell and bypass diode-activated) were chosen, based on relevance and representativeness. These categories cover a broad spectrum of fault characteristics and ensure sufficient sample sizes for meaningful statistical analysis. Less common defect types were excluded due to limited data, which would compromise the statistical reliability of the results. For each panel, a set of n images was taken, as shown in the following table. For each image of a panel, the identification of the panel’s dimensions, the calculation of the area in pixels, the identification of the defect area, the calculation of the defect area in pixels, and the transformation of the area into cm² based on the dimensional reference (panel dimensions: 99.2 × 164.4 cm) were performed, with results from

Table 1. Accuracy assessment of automatic defect area detection in PV panels compared to manual measurements.

Panel No.	Defect Type	No. of Images on Set (n)	Manual Detection Defect Area (cm2)		Automatic Detection Defect Area (cm2)		Absolute Error % absi
			Mean	Standard Deviation	Mean	Standard Deviation
1	Junction box	13	75	5.6	77	1.3	2.67%
2	Junction box	11	82	4.2	83	1.4	1.22%
3	Junction box	9	86	3.4	85	1.5	1.16%
4	Single-cell	8	219	4.4	221	2.3	0.91%
5	Single-cell	12	170	4.1	164	1.9	3.35%
6	Single-cell	9	182	3.8	179	2.1	1.65%
7	Multicell	8	321	4.6	325	2.2	1.25%
8	Multicell	14	422	5.1	415	2.3	1.66%
9	Multicell	8	437	3.1	438	2.1	0.23%
10	Bypass diode-activated	16	4511	10.1	4553	5.8	0.93%
11	Bypass diode-activated	8	4308	11.4	4289	6.7	0.44%
12	Bypass diode-activated	10	4527	14.8	4563	5.9	0.80%

.

For each panel, the average defect area and standard deviation were calculated based on the defect area values of each image in the panel set, using the formula:

$$\mathrm{S}\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{d}\mathrm{a}\mathrm{r}\mathrm{d} \mathrm{d}\mathrm{e}\mathrm{v}\mathrm{i}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}=\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{(x_i-\overline{x})}^2}{n-1}}}$$

(7)

where x_i is the value of the ith point in the dataset, $\overline{x}$ is the mean value of the dataset. and n is the number of data points in the set.

The identification of the edges of the panels and the defect area was carried out by two methods: a manual identification method and the automatic method proposed in this work. It can be seen from the table that the standard deviation of the proposed method is higher than the manual one (given that the manual operation time is large).

Therefore, the proposed method accurately identifies the presence of operating defects (manifested by an increase in local temperature) and determines the share of defects in the total surface of the panel (the total surface will be identified using mini-patterns).

4. Discussion

Given the values of the percentage absolute errors calculated in Table 1, the accuracy of the tested method is determined as:

$$\mathrm{a}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}=100\mathrm{\%}-{\displaystyle \frac{1}{12}}\sum _{i=1}^{12}{abs}_i=98.59\%$$

The mean absolute error is about 10.3 cm², which is very small relative to the large range of defected areas, especially in panels with severe defects. Regarding the repeatability of the mini-pattern algorithm, the results demonstrated consistent performance across multiple inspections. In terms of computational cost, the algorithm maintained low processing times and modest resource usage, making it suitable for deployment on standard computing hardware, including drone systems. The method also exhibited strong robustness to occlusion, effectively detecting defects even in the presence of partial visual obstructions, such as dust, shadows, or foliage. Notably, the algorithm operates without the need of labeled datasets, thereby eliminating the requirement for extensive training and annotation, which is typically a limitation in supervised learning-based approaches. This enhances the applicability in the field conditions where data may be limited or unavailable. A comparative analysis was conducted to evaluate the proposed algorithm against established PV fault detection methods.

Table 2. Comparative analysis of PV panel fault detection methods.

Method	Accuracy (%)	Repeatability	Computational Cost	Occlusion Robustness	Training Required
Ghost Convolution with BottleneckCSP and Tiny Target Prediction Head (GBH-YOLOv5) [22]	93.4	Moderate	High	Moderate	Yes
solAIr: Deep learning-based system for thermal images [23]	94.0	High	High	Moderate	Yes
Improved MobileNet-V3 for PV fault classification [24]	97.8	High	Moderate	Low	Yes
Deep neural network VGG19 with the Jellyfish Optimization Search Algorithm (JFOSA) [25]	98.34%	High	High	Moderate	Yes
Thermal image and inverter data analysis for fault detection [26]	89.5	Moderate	Low	Low	No
Mini-pattern algorithm	98.59	High	Moderate	High	No

summarizes the results across the key parameters discussed above.

Building on the current 2D image-based framework, the integration of 3D reconstruction and predictive analytics presents a feasible extension. Techniques such as UAV-based photogrammetry with advanced point cloud and geospatial analyses [27] and multi-view stereo [28] could enable the accurate 3D surface mapping of PV modules, enhancing defect localization and facilitating the detection of geometric anomalies like warping or panel displacement. This added spatial dimension would improve fault interpretation, especially in large or irregular installations.

Furthermore, incorporating predictive analytics based on historical inspection data and environmental factors could support early fault forecasting and degradation modeling. Such capabilities would enable condition-based maintenance and improve long-term operational planning. While these enhancements may introduce additional computational demands, advances in drone imaging, onboard processing, and cloud-based analytics make them increasingly practical for field deployment. Overall, these extensions would significantly broaden the diagnostic and decision-support capabilities of the proposed method.

5. Conclusions

The paper proposed a novel image processing methodology for the quantitative evaluation of thermal defects in photovoltaic panels, utilizing drone-acquired visible and infrared imagery. In contrast to conventional pixel-based techniques, the proposed method enables the dimensional measurement of hot spots in millimeters by applying an advanced mini-pattern recognition algorithm for panel boundary detection, geometric reconstruction, and image calibration.

The mini-pattern strategy significantly improves robustness under real-world disturbances such as shading, contamination, or partial occlusion by allowing the independent evaluation and partial matching of sub-patterns. This ensures accurate panel localization and scale factor determination, which form the basis for reliable thermal mapping and defect quantification.

Experimental validation on 12 PV panels, with four types of defects, demonstrated the automatic detection method outperformed manual measurements in terms of accuracy. The proposed method consistently estimated the physical surface area of thermal anomalies and their relative share compared to the entire panel area. Furthermore, it also exhibited high consistency across repeated observations, indicating enhanced stability and repeatability.

The proposed method may serve for establishing the actual physical size of the hot spot, the fault type, and the ratio of defective areas to the total panel surface, and for an estimation of the associated performance degradation under real-world operating conditions.

Due to its scalability, automation potential, and resilience to imaging inconsistencies, the technique is well suited for integration into drone-based PV inspection systems and large-scale solar farm monitoring. Future developments will focus on extending the framework to include 3D reconstruction and predictive analytics based on temporal thermal patterns.