Enhancing Solar Plant Efficiency: A Review of Vision-Based Monitoring and Fault Detection Techniques

Abstract

Over the last decades, environmental awareness has provoked scientific interest in green energy, produced, among others, from solar sources. However, for the efficient operation and longevity of green solar plants, regular inspection and maintenance are required. This work aims to review vision-based monitoring techniques for the fault detection of photovoltaic (PV) plants, i.e., solar panels. Practical implications of such systems include timely fault identification based on data-driven insights and problem resolution, resulting in enhanced energy outputs, extended lifetime spans for PV panels, cost savings, as well as safe and scalable inspections. Details regarding the main components of PV systems, operation principles and key non-destructive fault detection technologies are included. Advancements in unmanned aerial vehicles (UAVs), as well as in artificial intelligence (AI), machine learning (ML) and deep learning (DL) methods, offering enhanced monitoring opportunities, are in focus. A comparative analysis and an overall evaluation of state-of-the-art vision-based methods for detecting specific types of defects on PVs is conducted. The current performance and failures of vision-based algorithms for solar panel fault detection are identified, raising their capabilities, limitations and research gaps, towards effectively guiding future research. The results indicate that shading anomalies significantly impact the performance of PV units, while the top five fault detection methodologies, according to preset evaluation criteria, involve deep learning methods, such as CNNs and YOLO variations.

1. Introduction

Solar energy is an inexhaustible clean power source that is increasingly significant in both technological advancements and environmental sustainability. As climate changes and the depletion of traditional energy resources pose serious threats to our future, photovoltaic (PV) systems emerge as a pioneering solution. These systems harness the photovoltaic effect to convert solar radiation into electrical energy, providing a clean and limitless power source [1].

The journey of PV technology began in 1839 when French physicist Edmond Becquerel discovered the photovoltaic effect. Since then, research and technological advancements have led to the development of the first PV cells, which now offer impressive performance and applications across various needs [2].

PV systems are not just technical innovations; they are ecological necessities. Each kilowatt-hour of electricity produced by photovoltaic systems reduces carbon dioxide emissions, making them vital for environmental protection. Additionally, solar energy provides economic benefits, especially in sunny regions where investing in photovoltaics can result in significant savings and profits [3]. Therefore, it is essential that PV plants remain reliable and function effectively, referring to excellent installation, frequent maintenance, and high-quality inspections for faults. Faults may occur while manufacturing, transporting, assembling and operating PV panels. Faults need to be detected on time, so as to be repaired and prevent safety issues and energy losses.

Currently, fault detection methods have been in research focus, especially those that are not intrusive and can also provide details on the location of faults. Vision-based inspection has prevailed over other methods, due to the continuous advancements of camera technologies and graphics processing unit (GPU) architectures, as well as the progress of artificial intelligence (AI) and computer vision (CV) algorithms. The latter combination has the potential to detect, classify and localize faults, with low costs, and high reliability.

Contemporary challenges in vision-based PV fault detection systems include data overload issues due to the vast amount of produced data from various integrated sensors, creating barriers to the need for real-time fault detection. High-resolution images are needed to accurately detect faults; yet, their processing requires significant time and resources. Contemporary challenges also include the need for the improvement of fault detection accuracies by developing more advanced detection algorithms. The investigation of common faults, resulting mainly due to external environmental conditions, need to be identified so as to focus research on their effective detection.

To this end, this work aims to provide a review on AI vision-based solar panel inspection techniques, focusing on dealing with contemporary challenges. Many reviews have been recently published on the subject [1,3,4,5,6,7]. Yet, most of these review works focus on specific monitoring techniques and are not meant to be generic so as to cover multiple aspects of PV systems. These approaches aim to holistically summarize the key detection technologies, as well as machine learning and deep learning methodologies. The primary gaps we identified in these review works, and eventually aim to address in the proposed work, are the following:

• The lack of study on faults in relation to the performance of photovoltaic modules.
• The lack of evaluation of the proposed methodologies.

To this end, the contributions of this work that are not addressed cumulatively in other papers are summarized as follows:

• Presentation of PV systems’ fundamentals.
• Evolution of computer vision algorithms for PV fault detection within the decade.
• Summarization of all key monitoring techniques for PV fault detection, along with individual and comparative assessment of their capabilities and limitations.
• A focus on PV fault detection using AI-based computer vision, including machine learning and other pattern recognition methods, image processing techniques, and deep learning methods.
• Common faults detectable by CV algorithms in PV systems and how they affect the systems’ performance.
• Review of CV-based fault detection methodologies, cumulative performance tables and guidelines to select the appropriate one based on proposed criteria.

The rest of the paper is structured as follows. Section 2 presents the research methodology. Section 3 summarizes the fundamentals of PV systems. Section 4 presents the evolution of CV algorithms in PVs over the last decade. Capabilities and limitations of basic detection technologies are included in Section 5, focusing on fault detection using AI-based CV algorithms. Section 6 summarizes CV detected PV faults and related PV system performance. Section 7 provides an overview of CV-based PV fault detection methodologies and their evaluation. Finally, Section 8 and Section 9 discuss the research findings and conclude the paper, respectively.

2. Research Methodology

To provide a comprehensive review that manages to effectively address the identified gaps, we pose the following four research questions (RQ):

• RQ1: How have CV algorithms evolved in the context of fault detection in photovoltaic systems over the past decade?
• RQ2: What are the capabilities and limitations of key CV-based detection technologies concerning faults in photovoltaic systems?
• RQ3: What are the common faults that can be detected with CV in photovoltaic systems, and which of these significantly affect system performance?
• RQ4: What CV-based fault detection methodologies are identified in the literature, and how can the appropriate method be selected?

Figure 1 outlines the followed roadmap for addressing these four research questions. Research questions in this work are answered through a comprehensive review of identified articles relevant to the subject of interest. Figure 1 aims to illustrate the RQs and define the connection between them, as well as to indicate the exact order in which the questions are answered in the subsections entitled with the same name as each RQ, i.e., to align the RQs with specific subsections towards guiding the reader through the content logically.

The research of relevant articles in this work was conducted in the Scopus database within Article title, Abstract and Keywords, by using the query: “computer vision” AND “PV” AND “faults detection” OR “defects” OR “anomalies”. The search returned 52 articles ranging from 2014–2024.

To further refine the research results, the following eligibility criteria were applied:

• Language was limited to English.
• Subject area was limited to Engineering and Computer Science.
• Document types were limited to Conference papers and Articles.

By applying these limitations, 38 related documents were identified and used as sources in the following.

3. Fundamentals of PV Systems

In this section, the fundamentals of PV systems are summarized, including the main components of a PV system and PV operating principles.

3.1. Components

A PV system comprises the following main components, as illustrated in Figure 2:

• Photovoltaic Panels (PV Modules):

  a.

  Made up of photovoltaic cells that convert sunlight into electrical energy.

  b.

  Panels can be monocrystalline, polycrystalline or thin-film, each with unique performance characteristics and costs.

• Inverter:

  a.

  Converts the direct current (DC) produced by the photovoltaic panels into alternating current (AC) for use by electrical appliances or the grid.

  b.

  Different types of inverters, such as half-bridge and full-bridge, have various applications and features.

• Mounting System:

  a.

  Includes the supports and structures that hold the photovoltaic panels, either fixed or adjustable for solar tracking systems.

• Wiring and Connections:

  a.

  Essential for connecting the photovoltaic panels to the inverter and the power grid.

  b.

  Includes DC and AC cables, as well as grounding and lightning protection systems.

• Telemetry System:

  a.

  Ensures the monitoring and control of the photovoltaic system’s performance.

  b.

  May include wireless or wired connections for data transmission.

3.2. Operating Principles

The photovoltaic effect relies on semiconductor materials that convert solar radiation into electrical energy. When photons (packets of energy from sunlight) hit the photovoltaic cells, they excite the electrons in the semiconductor materials, causing an electric current to flow. This current is then collected and converted into usable electrical energy by the inverter.

The efficiency of a photovoltaic system depends on several factors, including the quality of materials, the intensity of solar radiation, and environmental conditions. Despite its limitations, photovoltaic technology continues to advance, offering more efficient and cost-effective solutions for generating clean energy.

In summary, photovoltaic systems represent an innovative and sustainable response to modern energy challenges, providing an environmentally friendly and economically viable solution. Continuous research and development in this field promise a bright future for solar energy and its role in creating a sustainable and clean environment.

4. Evolution of Computer Vision Algorithms in PVs over the Last Decade

This section provides a comprehensive overview of how computer vision algorithms have evolved in photovoltaic fault detection over the last decade (RQ1). Using Scopus, we explored over 60 relevant studies.

Table 1. Evolution of computer vision algorithms in PV fault detection over the last decade.

Period	Focus	Technologies/ Algorithms	Key Developments	Limitations
Early 2010s	Traditional Image Processing	Edge Detection, Segmentation [8]	Rule-based techniques, manual parameter tuning	Manual feature extraction, accuracy of fault detection
Mid 2010s	Introduction of Machine Learning	SVM [9], Random Forests [10]	Use of handcrafted features, initial use of feature classifiers	Feature engineering, scalability, complexity of faults, generalization
Late 2010s	Shift to Deep Learning	CNNs [11]	Automated feature learning, significant accuracy improvements	Benchmark datasets, training time, computational resources
Early 2020s	Integration of Multiple Modalities	CNNs with IR, Visible, and EL Imaging [12,13,14]	Enhanced fault detection under various conditions	Data overload, synchronization, fusion, computational burden, interpretability
Current Trend	Real-Time Detection and Adaptability	CNNs, YOLO, On-Site Processing [15]	Real-time processing, adaptability to environmental changes	Cost and scalability
Future Direction	Scalability and Economic Viability	Scalable DL models and Economic Analysis [16]	Scalable deep learning models incorporating economic analysis	Cybersecurity

includes the findings we have traced regarding the advancements and shifts in focus in this scientific field. More specifically, for each distinct period within the decade, the focus of research, the current state-of-the-art technologies and algorithms, as well as the key developments and limitations of each period are identified and summarized in the table.

In the early 2010s, traditional image processing techniques like edge detection and segmentation dominated the field [8]. These methods relied heavily on rule-based techniques and manual parameter adjustments. The latter posed limitations, since manual feature extraction can be time-consuming and may fail to capture the inherent characteristics of faults, leading to poor detection accuracies especially for complex faults. As the decade progressed, machine learning started to make its mark. Algorithms such as Support Vector Machine (SVM) [9] and Random Forest (RF) [10] appeared, employing handcrafted features and initial feature classifiers to improve detection accuracy. The limitation of feature engineering, feature extraction and selection, was still evident, while problem complexities could not efficiently be handled. Moreover, generalization issues were reported due to the varying environmental conditions and different PV systems, making algorithms difficult to generalize well. By the late 2010s, the field experienced a significant shift with the advent of deep learning, particularly Convolutional Neural Networks (CNNs) [11]. These models enabled automated feature learning, leading to substantial improvements in detection accuracy and efficiency. However, data requirements were excessive. Large benchmark datasets with annotated data were needed, while the training time of algorithms and the needs for computational resources began to augment. In the early 2020s, the focus shifted towards integrating multiple modalities, using CNNs in combination with infrared (IR), visible and electroluminescence (EL) imaging [12,13,14]. This integration resulted in enhanced fault detection capabilities under various conditions. Multiple data modalities posed many challenges, such as data overload, and data synchronization. Moreover, the need for sophisticated fusion techniques emerged. The processing of multiple data streams increased system complexity and computational burden. At the same time, the need for explainable deep learning algorithms started to show, revealing the research gap of interpretability of multimodal deep models. Currently, the trend is moving towards real-time detection and adaptability [17]. Advanced models like CNNs and YOLO are being employed for on-site processing, enabling real-time detection and adaptability to varying environmental conditions [15]. Interpretable AI algorithms, adaptability to new faults, and robustness to environmental changes were, therefore, in focus. Limitations included costs and scalability, since the size of PV installations tend to increase, producing vast amount of data and needing significant computational power.

Looking ahead, the future direction of research in this field is expected to focus on scalability and economic viability. Researchers are working on developing scalable deep learning models that can be economically viable and provide comprehensive fault detection solutions. A foreseen challenge is that of cybersecurity, since AI systems may be vulnerable to cyber and adversarial attacks, requiring robust AI models and secure communication protocols and encryption techniques to protect algorithms and data integrity.

Subsequently, by analyzing the content of the studied papers, and by using the functionalities of Scopus, Figure 3 was derived. Figure 3 depicts the key developments and research efforts, in terms of published works, over the past decade regarding the use of computer vision algorithms for detecting faults, defects and anomalies in PV systems, along with the key technological developments that marked the decade. The research was conducted within Article title, Abstract and Keywords, by using the query: “computer vision” AND “PV” AND “faults detection” OR “defects” OR “anomalies”. The search returned 52 articles ranging from 2014–2024. It is clear that the introduction of CNNs in PV fault detection in the late 2010s has brought a significant increase in the number of related documents, resulting an upward trend in the following years.

5. Capabilities and Limitations of Basic Detection Technologies

Various approaches are used for detecting faults and failures in PV cells and modules. These approaches are based on visual inspection, electrical measurements, electromagnetic radiation measurements and imaging techniques. In what follows, the key fault detection technologies for PV systems are reviewed, reporting their capabilities and limitations. Moreover, extended research in CV based fault detection technologies is conducted (RQ2).

5.1. Key Fault Detection Technologies for PV Systems

5.1.1. UAV-Based Inspection

The effectiveness of basic detection technologies has been tested in several preliminary laboratory experiments. These promising results have encouraged researchers to explore using these inspection techniques for large photovoltaic installations. The solution identified by field experts, which has become a standard in recent years, involves using specialized unmanned aerial vehicles (UAVs), such as drones, equipped with cameras for fault detection [18]. This method has two main advantages: (1) it is suitable for inspecting large photovoltaic fields as the entire installation can be surveyed in a few flights (depending on the size of the installation and the UAV’s battery life), and (2) it significantly reduces the cost and time required for analysis compared to traditional inspection techniques. The suitability of this solution has been evaluated through qualitative analysis in real photovoltaic fields to verify which anomalies are visible with moving cameras and to determine the impact of flight parameters on detection capabilities, thus testing the reliability of specialized UAVs [19,20,21,22,23,24].

Specifically, two types of architectures for anomaly detection can be identified: (1) real-time image analysis of photovoltaic modules during the UAV flight, and (2) offline image analysis, examining the video recorded during the inspection. Naturally, the first solution is faster and allows the operator to verify in real-time whether the photovoltaic modules have been inspected correctly; yet it requires careful software design to accelerate the image processing algorithm. Moreover, for real-time inspection, UAVs need to operate at certain heights and at specific time in the day so as to clearly capture details without sun reflections and facilitate processing. Moreover, note that UAVs can detect only visible external defects, while internal defects are not considered, e.g., non-functional diodes.

UAVs typically carry thermal cameras, visible light RGB cameras and photoluminescence cameras. Thermal cameras are used for thermographic analysis, while visible light cameras can store color images related to potential anomalies identified by the image processing algorithm. The idea is that images of defective photovoltaic modules, accompanied by geographical coordinates measured by the GPS sensor, can serve as a powerful and effective tool for automatic photovoltaic inspection [20]. However, the current accuracy of GPS sensors is in the order of several meters, whereas the required accuracy for the geographical placement of photovoltaic panels should be within a few centimeters. Therefore, with the current state of technology, computer vision algorithms cannot utilize the geographical coordinates calculated by GPS sensors to recognize and identify each photovoltaic panel but must provide a mechanism for tracking each unit and detecting potential anomalies in the image. The latter is a challenge being addressed by several techniques, as reviewed in the following.

5.1.2. Visual Inspection

Visual inspection is the fastest and most effective approach for identifying defects and faults in a PV module. This approach, however, is not suitable for modules exposed to weather conditions. Moreover, it must be carried out before and after the exposure of PV modules to mechanical, electrical, or environmental stresses. Various stress testing methods can evaluate modules indoor. Some common stress testing methods include moisture cooling cycles, thermal cycles, liquid heat testing, ultraviolet (UV) radiation testing, mechanical loads, hail impact, thermal stress application, etc. [25]. International Electrotechnical Commission (IEC) standards 61646 [26] and 61215 [27] require illumination of more than 1000 lux for visual testing, while defects visible to the naked eye are considered.

Common defects detected through visual inspection include delamination, yellowing and blistering on the front of the module; cracks, broken parts and discoloration of the reflective coating on the cells; burning and oxidation of metallization; bending, breaking, scratches and poor alignment of the module frames; delamination, yellowing, scratches, burning and blistering on the back of the module; corrosion, oxidation and loosening of the junction box; and fragility, disconnection and exposure of electrical components at connectors or wires [4,28].

Visible defects such as yellowing of the encapsulant have been identified as major causes of power loss [29,30]. Power losses can be measured by comparing existing I–V measurements (current–voltage characteristic) with values provided by the manufacturer. Modules with visible defects can be compared with a reference module (without defects) of the same characteristics to assess the impact of the defects on the module’s performance [29].

Several visual inspection studies have been conducted on PV modules exposed to different climates. For example, Bouaichi et al., in their study [29], found discoloration as a major cause of power loss for modules exposed to the Moroccan climate for two years. Discoloration was observed on modules above the junction box location. They found that the power difference between the module affected by discoloration and the reference module was directly related to the discoloration and indicated power loss due to discoloration. Kahoul et al. [30] also found encapsulant yellowing as a primary source of power loss for modules exposed to harsh climatic conditions (high summer temperatures, high radiation exceeding 1000 W/m² and sandstorms) for about 11 years. Additionally, cracks in the cells, degradation of the reflective coating, corrosion of the busbars, etc., were observed. Bouraiou et al. [31] found encapsulant yellowing and partial shading as primary sources of power loss for modules exposed to a Saharan environment in Algeria for about 12 years. Furthermore, delamination, corrosion, visible cell cracks, glass breakage, degradation of the reflective coating, etc., were observed.

Discoloration is a frequently occurring defect in PV modules operating in desert environments. Bouraiou et al. [32] found encapsulant discoloration in 608 (100%) of the units under study exposed to outdoor conditions in Algeria. Other observed defects include snail trail cracks, discoloration, delamination, corrosion, visible cell cracks, glass breakage, contamination, etc. Defects found with high rates included delamination and corrosion of the busbars. Visible cell cracks and glass scratches were also observed in some units.

5.1.3. I–V Curve Measurements

Measuring the current–voltage (I–V) curve is one of the main methods for characterizing solar cells [33]. According to [34], I–V measurement is considered the most comprehensive examination method. In such cases, measurements are first converted to standard test conditions (STC) and then compared with the values provided by the manufacturer. Indoor, controlled artificial light sources and temperature control systems are used, allowing for the maintenance of standard conditions. Studying the changes (deviations) in the current–voltage curve before and after faults can lead to the identification and investigation of module degradation. If the extent of the variation in the current–voltage characteristics is small, then it may be difficult to analyze faults, as minor failures in a module or cell do not significantly affect the current–voltage characteristics [35]. Therefore, it is challenging to detect minor failures using current–voltage measurements. Another disadvantage of this method is that the fault location cannot be pinpointed through current–voltage characteristics [36].

There are other diagnostic methods related to electrical measurements, categorized as electromagnetic tests [37]. Electromagnetic tests are used to identify internal degradation and related characteristics by analyzing changes in magnetic or electrical properties. These methods include DC parameter tests, AC parameter tests, the light beam-induced current method, the electron beam-induced current method and the superconducting quantum interference device technique. In DC testing, the modeling approach can be used to determine the DC parameters of solar cells [38] and the parameters can be analyzed graphically or theoretically. In AC testing, the AC parameters of solar cells are determined [39]. The light beam-induced current technique can map the light flow in the solar cell. The electron beam-induced current method can identify recombination areas, inhomogeneities and anomalies in the electrical characteristics of cells [40]. Combining the electron beam-induced current method with scanning acoustic microscopy can help identify electrical and morphological failures in solar cells [41]. Finally, the superconducting quantum interference device is an instrument used to measure changes in magnetic flux [42]. Additionally, other measurements such as resistance, current, voltage and magnetic induction can be performed [43] and used also for analyzing solar cells as they can measure the excitation current in cells and identify micro-cracks [44,45].

The final method to conclude this category is called differential current analysis, which is used to investigate the effect of non-uniform discoloration on photovoltaic cells [25]. Non-uniform discoloration in cells leads to uneven light transmission on the cells, resulting in electrical mismatches. Studying individual cells in the module can provide information on the extent of discoloration in each cell. For the non-destructive study of each cell, partial shading is used. In this process, each cell is partially shaded one by one, and the corresponding short-circuit current of the module is measured under standard test conditions. As the cells are connected in series, the current of the shaded cell limits the module current. This indicates the combined effect of discoloration and shading. Since shading is the same for all cells during the measurements, the variation in current in each cell provides information on the corresponding effect of discoloration. Partial shading of 50% can be maintained for all cells, and its effect is significantly greater than the effect of discoloration on any cell.

5.1.4. Infrared Thermography

Infrared thermography (IR) is a technique that involves measuring the surface temperature of PV modules. By using IR imaging, one can locate defects and assess their impact on power performance. Studies have shown that there is a correlation between cell power output and temperature variations in IR images. Specifically, the rate of power degradation is proportional to the temperature differences indicated by thermal images [46]. IR imaging involves capturing the infrared rays emitted by PV modules using thermal cameras. These cameras, known as infrared cameras, detect rays within the electromagnetic spectrum, between the visible range and microwaves, typically from 750 nm to 1 mm in wavelength. The thermal cameras commonly used in these applications operate in the 7 to 14 µm range, which falls within the mid-infrared region [47]. The thermal signal captured by the camera results from solar irradiation and local emission, following the Stefan–Boltzmann law:

$$\mathrm{P}=\mathsf{\epsilon }\times \mathsf{\sigma }\times \mathsf{A}\times {\mathsf{{\rm T}}}^4$$

(1)

where ε is the emissivity, σ is the Stefan–Boltzmann constant, A is the area, and T is the temperature.

There are four main types of thermographic measurements: steady-state, lock-in thermography, induction thermography and pulse thermography:

Steady-state thermography is the most commonly used technique and allows for analysis during module operation. It involves capturing thermal maps of PV modules, with abnormal temperature areas indicating potential defects. This method can be performed in outdoor (sunlight) or indoor (dark) environments. Outdoor measurements, also called outdoor or illuminated thermography, are taken when ambient temperatures are low and wind speed is normal [48]. Indoor measurements, known as indoor or dark thermography, require disconnecting the modules and applying a current comparable to the module’s short-circuit current [47]. The setups for indoor and outdoor thermography [49] are illustrated in Figure 4.

Lock-in thermography involves exciting samples at a controlled frequency, periodically stimulating cells to reduce noise and enhance the signal-to-noise ratio (SNR), thereby detecting weaker heating sources. This technique has less thermal impact on cells and can be performed in dark or illuminated conditions [50,51,52]. Induction thermography, also known as pulsed current method, involves inducing currents in materials using electromagnetic waves, which generate heat detectable by thermal cameras. This technique can reveal defects through variations in thermal diffusion [53,54,55]. Pulse thermography uses an external heating source, like a flash lamp, to create a dynamic heat flow through the module. The surface temperature rises uniformly and a high-resolution thermal camera captures images to detect defects like bubbles and electrical connections. Various defects visible in IR images [49] are shown in Figure 5, such as cell-to-cell connection failure, cracked cell, cracks isolating parts of the cell, high-resistance solder bonds, local bypass junctions, high current density in busbars, glass breakage and cell damage in external setup.

Considerations such as emissivity settings, solar irradiation, shading effects, connectors and support structure impact on temperature patterns are crucial during IR imaging [56]. Information on the thermal camera’s distance from modules is also essential as absorption by gases and water vapor in the air can affect results. Glass reflection can also pose a problem, creating measurement errors up to 15 °C depending on cloud cover during imaging. Adjusting the camera angle can help reduce reflection issues. A minimum solar irradiation of 500 W/m² is recommended for imaging [56].

Modules in hot climates tend to show more damage when hot spots are present compared to those in non-hot zones. Cracked cells with hot spots exhibit higher degradation than those without hot spots, indicating a significant impact on cell temperature and power losses [57].

5.1.5. Electroluminescence Imaging

Electroluminescence (EL) imaging involves applying a current comparable to the module’s short-circuit current (ISC) in the forward direction, causing cells to emit EL radiation due to electron-hole recombination [58,59]. These EL radiations are usually detected by charge-coupled device (CCD) cameras due to their relatively low cost. The emitted radiations lie in the near-infrared range and can be effectively captured by InGaAs (indium gallium arsenide) sensors, although these are more expensive. Modified digital RGB cameras can also be used by removing the IR filter to detect near-infrared emissions [60]. The process is carried out in a dark environment, where defects appear as dark areas or spots and cracks show up as dark lines in EL images [61]. Commonly detected defects include cracks, material defects, finger interruptions, etc. Studies have used EL images to investigate potential-induced degradation in PV modules [62]. EL images can sometimes have random dark spots or areas, making defect recognition challenging [63]. Neighboring pixels can be combined to improve the signal-to-noise ratio and images can be processed to remove noise and erroneous pixels. High-resolution imaging may require capturing individual cell images and stitching them together for a complete module image.

This technique is considered quick, efficient and accurate for indoor defect detection. An EL imaging experimental setup from the work of Akram et al. [11] is shown in Figure 6. From the same research work, typical crack types and other defects in EL images are shown in Figure 7, more specifically parallel to busbar, +45°, −45°, several directions, dendritic/branched, deep cracks isolating cell parts, cross line, perpendicular to busbar, finger failure, silicon material defect, contact forming failure and finger failure along cracks.

The average pixel intensity of an EL image of a cell is directly related to the module’s maximum power output per cell area [62]. This relationship helps determine the module’s degradation level. EL emission intensity correlates with the applied voltage level and series resistance losses appear as low-intensity areas in EL images. Defective regions can be identified by comparing EL images under different polarizations [64].

5.1.6. Photoluminescence Imaging

Photoluminescence (PL) imaging is another effective approach for detecting defects in PV modules. In this method, a sample is excited by light radiation from a laser source, causing it to emit PL radiation, which is detected by a cooled CCD sensor [65]. The emitted radiations fall in the near-infrared range and this technique can be used to investigate silicon wafers, rods, layers and cells. A typical PL imaging setup is shown in Figure 8.

PL imaging can also be performed using optical filtering and current modulation. In this method, the current of an individual cell in a series string is varied between normal operating and open-circuit points by deliberate shading using an LED. The entire string operates under the same conditions, achieving high-quality PL images. The individual cell under modulation is termed the control cell and the rest are test cells. The obtained PL image is compared with the EL image to identify cracks, high-disorder areas and poorly performing cells [66]. This enhanced PL imaging setup and the resulting image are shown in Figure 9.

Spectrally and spatially resolved PL imaging focuses on a line rather than the whole sample. The line signal passes through a diffraction slit before the focusing lenses and is spectrally separated [66]. This signal is detected by a CCD chip and spatial resolution is achieved by moving the sample or camera. This method detects cracks and surface contamination in cells [67].

5.1.7. Ultraviolet Fluorescence Method

The ultraviolet (UV) fluorescence method initially investigated discoloration in PV modules. Discoloration occurs due to environmental and other factors. When the encapsulant material (EVA) in PV modules is exposed to sunlight, particularly UV light, its molecules break down and form chromophores [28]. Chromophores are functional groups or atoms in a compound responsible for its luminescent properties.

In UV-based detection, a UV light source excites the chromophores in the encapsulant, causing them to emit fluorescence [68]. Specific types and patterns of fluorescence are formed by moisture, temperature and radiation pressure [69]. Mechanical failures like glass breakage and cell cracks significantly affect fluorescence quenching. The experiment is conducted in darkness and it is recommended to expose the module to sunlight before imaging. Longer exposure results in intense luminescence emission and a 30-s exposure time is advised for good fluorescence images [70]. Emitted rays are then imaged by a camera, providing information about cell cracks. The emitted fluorescent light has a wavelength in the 400–800 nm range. Other failures detectable by UV fluorescence imaging include isolated cell parts and disconnected cell interconnections [71]. An experimental UV fluorescence imaging setup is shown in Figure 10, including a PV module, a CCD camera, a highpass filter and a UV light source.

The emitted fluorescent light lies in the visible spectrum and a digital camera is used for imaging. Typically, filters are employed to block UV rays. The information provided by fluorescence imaging is similar to that of EL imaging [72]. Fluorescence inspection is also possible outdoors while the module is operational, without the need for disconnection. It can identify hot spots, cracks and cell mismatches and serves as a potential alternative to EL and thermography with fewer practical constraints [71]. For a detailed analysis of a material’s fluorescent characteristics, fluorescence spectroscopy can be performed.

5.1.8. Spectroscopy

Spectroscopy involves measuring and studying spectra produced by the interaction of matter with radiation. A spectroscopic device, called a spectrometer, measures electromagnetic radiation at specific wavelengths. Raman spectroscopy is one technique for obtaining detailed information about a sample’s chemical structure, molecular interactions and crystallinity. When light scatters off the material’s structure, the Raman effect occurs. Most incident light scatters elastically (Rayleigh scattering) at the source’s wavelength, while a small portion scatters inelastically, resulting in a wavelength shift. This shift characterizes the material’s composition. Raman bands shift to higher or lower wavelengths from the source wavelength, depending on the material’s properties. The intensity of the Raman effect is determined by the source wavelength, material concentration and sample dispersion properties.

A Raman spectrometer and hyper Raman head measure Raman spectroscopy maps to detect snail trails in PV modules [73]. The hyper Raman head filters unwanted laser sidebands and Rayleigh scatter. A CCD sensor with a spectrometer provides high-intensity signals. To map snail trails, point measurements are converted into Raman maps, which can also produce fluorescence images. A Gaussian filter may be applied to the Raman maps. Fluorescence intensity is lower along cracks and silver bands, while high intensities are observed in non-cracked areas. Dark and bright regions in fluorescence images help identify snail trails. Such a measurement setup is shown in Figure 11 [73].

Fluorescence spectroscopy can also study modules exposed to aging [70]. A UV light source excites the PV modules, causing them to emit fluorescent light, detected by a spectrometer via optical fiber. UV fluorescence spectra can be measured this way. Modules emit high fluorescence intensities before aging, which decrease after degradation.

Fourier transform infrared spectroscopy (FTIR) can analyze backsheet degradation in modules [70]. Interaction between mid-infrared radiation and matter excites molecular vibrations and the absorption wavelengths appear in infrared spectra, revealing molecular structure. Typically, an attenuated total reflection (ATR) mechanism propagates an evanescent wave through the module.

Raman spectroscopy also measures thermomechanical stresses in PV modules during manufacturing [74]. Stresses in cells before and after soldering and during lamination can be measured. The Raman effect defines the inelastic scattering of photons (called phonons as they create vibrations in cells) by matter. The Raman peak position depends on photon energy levels, reflecting lattice structure and is material-specific. Thus, any Raman peak change relates to mechanical stress-induced deformations. The Raman peak shift can be converted to stress using a linear conversion factor. The Gaussian function defines peak fluctuations, requiring numerous measurements for reliable results [74].

5.1.9. Electromagnetic Induction-Based Measurements

Recently, some existing methods, like thermography and electroluminescence, have been modified through electromagnetic (EM) radiation induction.

An IR thermography approach based on EM induction can detect defects in PV modules and cells. This approach provides quantitative evaluation. An induction coil induces EM currents in the cell, generating heat detected by thermal cameras. This process involves three steps: EM-based heating, heat conduction and IR emission. The setup includes a defective solar cell, induction heater, signal generator, thermal camera, induction coil, power supply and computer. The coil is placed 5 cm above the cell, generating high-frequency AC signals. This method, applicable in pulsed and lock-in forms, can detect cracks, delamination defects, fatigue and micro-defects [75]. Figure 12a illustrates an active electromagnetic induction infrared thermography (EIIT) system for PV cells from the literature [75].

Electrothermography, a modified indoor thermography approach based on EM induction, also enhances defect detection capability. Similarly, modified EL imaging based on EM induction improves defect detection. Various defects like broken grid lines, scratches, hidden cracks, surface impurities, etc., can be identified using these techniques. EM induction significantly enhances defect detection in both thermography and EL. Additionally, merging EL and IR images obtained from these enhanced techniques provides more information. Image merging combines sparse vectors from electrothermography and EL images using L1 normalization. Yang et al. [76] compare sparse representation fusion results with curvelet, wavelet, dual-tree complex wavelet and contourlet transforms, evaluating performance based on five metrics: root mean square error, correlation coefficient, peak signal-to-noise ratio, mutual information and structural similarity index. Sparse representation outperforms other algorithms. The used EM induction-based thermography and EL imaging setup is shown in Figure 12b.

5.1.10. Capabilities and Limitations

After reviewing the key detection technologies individually, this section aims to identify their capabilities and limitations by examining fault detection technologies in combination, aiming to answer RQ2. To achieve this, we utilize two comprehensive tables (

Table 2. Cumulative table of limitations and capabilities of basic PV fault detection technologies. CV based fault detection technologies are marked in bold.

Technology	Description	Capabilities	Limitations
UAV-based inspection	Specialized UAVs equipped with cameras for fault detection fly over PV farms.	Suitable for inspecting large photovoltaic fields Reduces the cost and time required for analysis com-pared to traditional inspection techniques	Limited accuracy of GPS sensors Need to track each unit Only defects visible from distance can be detected Not suitable for real-time detection
Visual Inspection	Defects detected with naked eye, such as delamination, browning, yellowing, corrosion, bending, bubbling and degradation of the anti-reflective coating.	Quick and efficient No instrumentation required	Cannot detect non-visible defects Not feasible for large-scale outdoor applications
I-V Curve Analysis	A primary approach for characterizing silicon cells. Typically combined with other methods for detailed information. Changes in the I-V curve lead to identification of PV module degradation.	Low-cost methodology Easy measurements Can be used for quantitative calculations	Cannot pinpoint the exact location of defects May be ineffective with minor variations Contact method requiring instruments
TG/IR Imaging	A method that measures the surface temperature of PV modules. Infrared rays emitted by the modules are captured by thermal cameras. Various types suitable for different applications.	Suitable for large-scale outdoor applications Easily detects hot spots Provides quantitative measurements High-resolution images Non-destructive Can detect areas of internal short circuits	Difficult to precisely locate the defect Expensive thermal cameras Long measurement time with lock-in IR method Thermal blur issues Indoor IR requires external power source Micro-crack damage is not fully represented
EL Imaging	Captures electroluminescence radiation emitted by cells due to electron-hole recombination. This radiation is in the near-infrared spectrum.	Primarily for detecting micro-cracks and edge interruptions Fast, efficient and accurate for indoor use Non-destructive Can be performed with a modified digital camera	Random dark spots/lines/areas in the background due to crystallographic defects Requires more experience and expertise Requires external power source Mainly for indoor use Induction heating issues blur interior areas
PL Imaging	The sample is stimulated with light radiation/laser source and luminescence radiation is emitted near-infrared region.	Fast Non-destructive High spatial resolution Can detect cracks	Requires an excitation source Branched areas appear quite blurred
UV-F Imaging	Uses an ultraviolet light source to stimulate the luminescent pigments in the encapsulating material. This stimulation leads to the emission of fluorescence luminescence The emitted rays are then imaged by a camera.	Easy detection of snail trails Easy detection of discolorations Can detect cracks Fluorescent light is in the visible range, so a digital camera can be used Non-destructive	Requires long exposure time for good fluorescence image Requires light source for stimulation Fluorescence effect develops in modules after prolonged outdoor use Cannot detect PID Shorted or open bypass diode is not detectable
Spectroscopy	Measuring and studying spectra produced by the interaction with radiation.	Highly sensitive Can differentiate between different fault types	Costly equipment Complexity of spectral data Affected by external conditions
EM induction	Identifies variations in electrical properties caused by faults	Scalability for large systems monitoring Fast scanning	Complex interpretations Environmental interference

and

Table 3. Comparative table of limitations and capabilities of basic CV-based PV fault detection technologies.

Ref.	Technologies	Findings
[77]	Indoor vs. Outdoor TG/IR	External IR thermography results show relatively fewer or no defects in PV modules. Conversely, internal IR thermography images depict defects more clearly. Possible reasons for differences include absorption of radiation by other parts such as the backsheet, high heat dissipation rate, abrupt environmental changes causing thermal instability and minor defects of negligible impact.
[78]	UV-F vs. EL	Cracks in EL images clearly correlate with dark areas in UV-F images. However, due to darkness around cell edges, cracks along the edges are not detectable in UV-F images. The marble pattern in EL images caused by crystallographic defects in polycrystalline silicon makes crack detection in EL images harder compared to UV-F. UV-F better illustrates areas typically hotter during operation.
[79,80,81,82,83]	TG/IR Imaging vs. EL Imaging	EL Imaging	IR Imaging:
		Advantages
		High resolution Direct measurement (non-contact) recognizable defects: defective laser cut, shorted bypass circuits, disconnected cell areas, short circuits broken cells and layer defects	Recognizable defects: different thermal behavior, short circuits, hot spots, moisture, shading, incompatibilities, installation failures, etc.
		Disadvantages
		Origin of defect not recognizable Hard to determine defect impact on cell/module performance Normal-looking EL images can reveal high-temperature areas in IR images because both techniques capture different physical properties	Not all defects cause temperature rise High-temperature areas are not always defect sources Difficult to pinpoint exact defect location in numerous small spots Requires electrical interface Cannot distinguish between weak and high series resistance
[84,85,86]	Visual Inspection vs. TG/IR vs. UV-F	Hot-spots easily detectable with IR, but EL and UV-F do not clearly detect them Hot-spots above 120 °C easily visible with visual inspection, appearing as dark black or brown area in RGB images Cell cracks not clearly detected with IR thermography but evident in EL images. UV-F shows similar crack patterns, but formation can take weeks Snail trails are easily detected with UV-F and EL Potential Induced Degradation (PID) faults are detected with IR thermography, EL images, but not with UV-F Shorted/opened bypass diodes detectable by all methods except UV-F and partially by visual inspection Common visible faults include discoloration, broken glass and backsheet tearing

) that detail the strengths and limitations of the different basic detection technologies and ultimately how they relate to various faults in photovoltaic systems, based on an extensive literature analysis.

Besides the individual application of each basic detection technology, various studies examine the comparative application of multiple CV-based technologies on PV modules. The findings of these comparative applications are summarized in Table 3.

5.2. Fault Detection Using AI-Based Computer Vision

Computer vision is emerging as a promising technology to replace traditional monitoring systems in the PV field, addressing the world’s growing energy needs due to the exponential population increase. Integrating AI into PV energy systems is becoming a hot topic, as AI will play a crucial role in meeting future energy demands.

Computer vision systems are automated image recognition systems used in the PV field to introduce intelligent behavior in computers, cameras, smartphones, machines and drones. In other words, this technology involves making a computer, camera, or machine to “see” and behave intelligently. It uses a camera with a sensor and a computing system to capture and detect objects instead of the human eye and brain. The information is then used by control systems for further processing and action, enabling systems to exhibit human-like intelligence in laboratories, industries and working fields.

Various types of sensors, such as RGB, EL, IR, etc., are used in PV for ground and aerial imaging. These images are learned and recognized using machine/deep learning algorithms and statistical methods, leading to improvements in quality, accuracy, timely response, energy and labor savings, increased energy productivity, risk reduction and higher outcomes.

Automated fault detection methods based on computer vision and AI in PV modules, cells and arrays involve introducing intelligent behavior in machines or computers to detect failures in cell images, assess silicon wafer quality, predict faults, etc. In other words, the monitoring and fault detection process in PV systems is automated using machine learning algorithms, image processing techniques and deep learning methods, as documented in the literature. Deep learning has shown excellent performance in this field. For example, replacing EL image analysis of PV cells or modules with an automated image classification and detection system, automatic edge detection in infrared (IR) images of PV modules, automated defect detection and localization in RGB images of PV modules, silicon wafer quality assessment in PL images, crack segmentation in EL images, automatic fault detection in grid-connected PV installations and machine vision-based detection in silicon wafers.

In these examples, human intelligence is exhibited by computers or machines. Therefore, such methods are categorized as automated or AI-based methods, whether applied to images (IR, PL, RGB, EL, etc.) or other data [87]. AI-based methods are also used for prediction, modeling, predictive analysis and other purposes [88,89,90]. These computer vision or AI-based approaches use different types of machine learning and statistical algorithms and image processing techniques. Commonly used algorithms include SVM, ANN, CNN, etc., for regression and classification problems. Image processing techniques like segmentation, Canny edge detection, filtering, contouring, color quantification, morphological operations and data augmentation are also widely used in automated applications. Feature extraction and selection techniques followed by classification are also employed. Currently, AI-based automated approaches are marginally used in PV monitoring. However, with increasing energy demand and continuous installation of large PV plants, automated methods will become essential to meet various requirements. For large-scale PV installations, automatic inspection using unmanned aerial vehicles (UAVs) and computer vision algorithms is commonly used.

5.2.1. Machine Learning and Other Pattern Recognition Methods

Machine learning algorithms, like SVMs, can be used for automatic classification of normal and defective PV cell/module images. Demant et al. [91] classified cracked and normal photoluminescence images using the SVM algorithm. In [92], crack patterns were defined using the gradient location and orientation histogram (GLOH). A limitation of the method is that it is sensitive to region errors, which can be corrected, but will result in a very time-consuming algorithm implementation.

Feature extraction followed by a classifier can also be used for automatic classification of normal and defective PV cell/module images. Kato in his study [93] used the SVM classifier on extracted image features to automatically classify failure modes in RGB images of PV cells caused by impact with metal conductors and various forces. These failure modes were detected from recorded video after image extraction. In this study, HOG [94] and SURF [95] features were extracted from images and fed into the SVM classifier.

Additionally, pattern recognition followed by the SVM algorithm is used for classification. A study by Demant et al. [96] performed automatic crack detection in photoluminescence and infrared images of PV modules using pattern recognition based on local descriptors and the SVM algorithm for classification. The radial basis function kernel was used in the SVM application.

The Fourier image reconstruction technique was also used for automatic detection of defective solar cells [63]. Defects appear as bar or line-shaped objects in this approach. Experimental results showed that the proposed approach was effective for detecting a few defects. This method had some complications in detecting defects with more complex shapes and took 0.29 s to test one cell. Additionally, only a few defects like finger interruptions, micro-cracks, etc., could be detected.

Crack saliency maps generated by the evidence filtering process were used for automatic crack detection in [97]. Crack saliency maps of EL images of PV cells were generated by evidence filtering and the local threshold process processed the obtained saliency map for automatic crack detection. Crack extraction was based on segmentation. Then, the minimum spanning tree connected crack fragments, and finally, cracks were detected by skeleton extraction. A drawback of the method was that the segmented crack may miss a part when morphological operations were applied to remove non-crack pixels, affecting the performance of the detection algorithm especially when curvilinear crystal grains were in the background.

Independent component analysis (ICA) was also used for automatic PV defect detection [98]. The ICA reconstruction process was as follows: (1) a defect-free image was selected as a training image, centered and whitened, (2) independent components and the unmixing matrix were computed using the fast ICA algorithm, (3) independent components were sorted in descending order and rearranged, (4) the number of components was selected, (5) the unmixing matrix was reshaped, and the image was reconstructed, (6) the inspection image was centered and whitened, (7) the inspection image was reconstructed, (8) the reconstructed source image was binarized. A drawback of this method was that finger interruptions in cells were treated the same as small and deep cracks.

In general, the limitations of machine learning algorithms, as already mentioned in Section 4, are due to the need for representative features for each fault type. The latter process can be time-consuming and could prove insufficient to capture the characteristics of each fault. Traditional machine learning methods fail to generalize well in case of new unseen data, which is a limitation in case of PV faults, where input images can be under various and different environmental conditions. Moreover, large datasets and high-resolution images are hard to feasibly handle and process in real-time.

Referenced methods of this section are further comparatively evaluated in Section 7, based on preset performance criteria.

5.2.2. Image Processing Techniques

Image processing techniques can also be used for automatic fault detection in PV modules [99]. Several image processing schemes were used in existing studies as described in the following.

In [100], the authors obtained infrared images of a large PV system from a UAV. The images were first converted to grayscale and then segmented. Morphological operations followed by Canny edge detection were applied. This research also showed the results of other image processing schemes. The main limitation of the method was reported to be the existence of shadows in images that obstructed the detection of some hot spots.

In [101], the authors used different image processing methods for automatic detection of visible defects like dust and snail trails. Images of a large PV system were acquired from a UAV and converted from RGB to single-channel models (red, green and blue channels), reducing computational complexity. After channel separation, image filtering reduced noise. The first derivative of the Gaussian function (FODG) was used for image filtering. An edge detection algorithm was then applied to detect defects. However, the system could perform efficiently only for UAV flights of specific height, under specific sun-angle and only in optimal weather conditions.

An image processing scheme involving segmentation combined with the Canny edge detection method can be used to identify hot spots in infrared images of PV modules [102,103]. Both works reported as limitation the fact that Canny edge, in case of any specular object present in the background, caused grey-level variations that were false assumed as hot spots. Segmentation is also used to detect defects in infrared module images [104], yet was not efficiently performing in case of whole modules with several defects. Another research work [105] used different image processing approaches and statistical methods for automatic evaluation of aerial infrared images of modules. Segmentation combined with morphological methods was also used for automatic detection of hot spots in modules [106], yet, was requiring further research to detect specific defects other than hot spots.

Filtering operations combined with segmentation techniques can also be used for automatic defect detection [99,107]. An image processing scheme involving filtering, color quantization and Canny edge detection can detect defects of varying severity in infrared images of PV modules was reported in [47]. This scheme could identify normally functioning, slightly defective and severely defective areas in infrared images of PV modules.

Finally, the vesselness-based algorithm for segmentation can also be used for the automatic fault detection scheme. This method was used in [108] for EL images. It has some limitations, such as only considering cracks longer than 20 mm, and disregarding other defect types and cracks shorter than 20 mm.

Same as in traditional machine learning algorithms for PV fault detection, the limitations of image processing methods, as already mentioned in Section 4, are due to the need for exhausting manual feature extraction. Such methods are sensitive to noise and environmental variations, leading to poor adaptability, requiring fine-tuning for each specific case, and subsequently result to poor detection performances. Scalability to cover large PV farms, and complexity of large datasets, are also evident.

Referenced methods of this section are further comparatively evaluated in Section 7, based on preset performance criteria.

5.2.3. Deep Learning Methods

Besides the aforementioned approaches based on machine learning, image processing and other traditional pattern recognition methods, deep learning algorithms have also been used for PV fault detection. Studies [109,110,111,112] utilized deep learning to detect visible defects. Another study [113] employed CNNs for quality control and process monitoring during solar cell production, limited by the low resolution of the used camera.

The Mask FCNN network was used in [114] for pollution type prediction and localization in solar modules, also predicting power loss. Such deep learning-based methods primarily focus on visible defect detection. Recently, some studies have applied deep learning to EL and IR images. In their study [114], Mehta and Azad, used a deep learning-based approach for automatic defect classification in EL images. This study also used the SVM algorithm with feature extraction techniques like KAZE, SIFT, HOG and SURF.

However, the best results were obtained by using CNNs with transfer learning. In [115,116], the authors suggested CNN approaches, utilizing a publicly available dataset of EL images of solar cells and then implemented a transfer learning approach using the VGG network. However, errors in the labelled dataset, especially at the crack edges, seem to greatly affect the performance of the detection model. A lightweight CNN architecture can also be used for defect detection in EL images of PV cells. In [11], the authors used a lightweight CNN architecture and adopted generalization strategies to achieve good performance using ordinary hardware resources. They used various data augmentation strategies to address data scarcity and maintained real-time prediction speed. Different defect types in EL cell images were examined, yet the method was tested only in the lab.

Both standalone deep learning and transfer deep learning can successfully detect defects in IR images of PV modules. The study of Akram et al. [49] gathered an IR image dataset after experiments on normal and defective modules. For standalone learning, they used a lightweight CNN trained from scratch. For transfer learning, they used model development techniques, where a base model pre-trained on another EL image dataset, transferred knowledge to the target model trained on the IR image dataset. Transfer learning with model development performed relatively better. They also discussed various defect types in IR solar panel images. Classification errors were observed for defects having a limited number of images in the dataset.

YOLO networks are a hot topic in recent research on fault detection in PV modules. Specifically, the authors in [117] used YOLO for detecting hot spots in IR images of PV modules. This approach employed bypass connections to concatenate features extracted from initial layers with refined features from later layers. This method segmented PV modules from images and detected hot spots. Experimental results indicated the robustness of the proposed method, achieving real-time speed without extensive setup.

VGG and MobileNet networks were also used to detect and classify defects in IR images of PV modules [118,119]. Collected IR images of defective and non-defective modules were also classified based on feature extraction. SIFT and dense SIFT algorithms extracted features, followed by SVM classification. For SVM, polynomial and radial basis function kernels were used, with the polynomial kernel yielding better results. Limitations such as the complexity of the model and computational resources needed to be considered towards identifying the best performing model for a given task.

Faster R-CNN was used to detect hot spots in thermal IR images of PV modules [12]. Pre-trained model weights were improved on the IR image dataset for this specific task. Additionally, an image processing scheme involving Hough line transformation and Canny edge detection was used to detect hot spots. The Faster R-CNN approach achieved excellent results but has a high computational cost, which is challenging especially for UAVs due to their limited memory and GPU abilities.

Buratti et al. [120] used transfer learning (AlexNet, ResNet, SqueezeNet and VGGNet) to extract features from EL images of PV cells and classify defective cells with about 96% accuracy. SVM regression predicted performance using extracted features. Training data included images of monocrystalline silicon cells and corresponding I-V parameters. The dataset included busbar-free cells, cells with 3 busbars and cells with 5 busbars. The system was validated on full-circuit cells with 9 busbars and half-cells, showing successful results.

These methods and applications highlight the potential of deep learning in detecting and classifying defects in PV systems, emphasizing their effectiveness and significance in the field.

In general, the limitations of deep learning algorithms, as already mentioned in Section 4, are due to the need for large amounts of labeled and diverse input data. Data acquisition and annotation could be exhausting. Moreover, deep models require powerful GPUs especially for real-time applications. Explainability of deep models is an additional challenge, closely related to the trust on their outputs, especially for stakeholders who want to invest in this technology.

Referenced methods of this section are further comparatively evaluated in Section 7, based on preset performance criteria.

6. CV Detectable Faults and Related Performance of PV Systems

Photovoltaic modules (PVMs) can experience various types of failures, often covered by the system’s warranty if they happen under normal operating conditions [93]. When these systems fail, it can lead to safety risks, reduced efficiency, decreased power availability and compromised system reliability. Common issues found in PVMs include discoloration, cracks, snail trails, damage to the reflective coating, bubbles, pollution, oxidation of distribution bars, corrosion and breakdown of encapsulation on cells and interconnections. Additionally, problems like loss of adhesion of the back sheet are also noted [28,121,122,123,124,125,126,127,128].

There are various strategies to detect these failures, focusing on issues like encapsulation, unit corrosion, cell cracks and problems with the photovoltaic inverter [128]. Failures in PVMs are generally categorized into two types: permanent and temporary. Permanent issues include delamination, bubbles, yellowing, scratches and burned cells, typically requiring the replacement of the defective modules. Temporary problems, like partial shading, dust accumulation, dirt and snow on the PVM, can usually be resolved by the users without needing to replace the modules. Failures can arise from both external and internal causes, both of which can decrease the system’s output power, efficiency and reliability.

The main types of failures that can occur in a PVM are summarized in

Table 4. Summary of fault types, effected elements, causes and effects on system performance.

Ref.	Fault Type	Affected Element	Causes		Effects
			External	Internal
[36,125,129,130,131,132,133]	Hot Spot Faults (HSF)	PV Cells, PV Modules	Covered by: Dust, Snow, Shadow Different classes of PV Modules or technology	Fragmentation of cells Current mismatch between cells High-resistance or “cold” solder points Aging and degradation of solar cells	Destruction of PV panels Open circuits Reduced efficiency Reliability issues
[93,134,135,136]	Diode Faults (DF)	Bypass Diode (BpD), Blocking Diode (BkD)	Partially shaded cells Overheating	-	Destruction of diodes Short circuited diode Open circuited or Isolated diode
[137,138]	Junction Box Faults (JB)	Junction Box	-	Wear Corrosion Loosen connections Oxidation	Destruction of PV cells Fire hazard Reduce efficiency Reliability issues
[4,28,123,139]	PV module faults (PVMF), PV array fault (PVAF)	PV Modules	Glass breakage of PV Modules frameless caused by the clamps Connector failure Ground Isolation Encapsulation Installation wiring mistake	Corrosion of solar cells Manufacturing defects Delaminated Bubbles effect Yellowing Scratches Burnt solar cells	Isolated modules Short circuit between modules Current leakage Destruction PV modules Reduce efficiency Reliability issues Reduced output power
[140,141,142,143]	Ground Faults (GF)	PV Array, PV String	-	Insulation failure of cables Wire ground fault Ground fault inside solar panel cable insulation during installation Ground fault inside solar panel from bad sealing Insulation destruction of cables Short circuit inside the solar panel junction box	Fire hazard
[144,145]	Arc Faults (AF)	PV Modules	-	Wire short break Two wires with potencies different are place near one other Bad soldering connection Leakage inside to solar panel from mechanical damage Wildlife junction box Loosening of screws	Destruction of PV panels Fire hazard
[146]	Line to line faults (LLF)	PV Array	-	Low resistance between two points with potencies different Insulation failure of cables Short circuit between wires Insulation failure between string connectors Mechanical stress	Destruction of PV panels Wires damage Fire hazard

. Moreover, their effects on the systems’ performance are also included in the table, along with the affected elements and the causes, based on the examined literature (RQ3). More details regarding failures and their relation to performance can be found in the following subsections.

6.1. Hot Spot Faults

Hot spot (HS) faults can happen when some cells in a photovoltaic (PV) panel have differing current–voltage (I–V) characteristics [125]. This typically occurs due to inconsistencies in manufacturing, which lead to high-resistance points or poor solder joints [125]. Dust and dirt buildup on the panels [129,130,147,148] (Figure 13b), aging of the cells, incomplete insulation at the edges with transparent materials, manufacturing tolerances and uneven sunlight exposure can also contribute to these issues. Partial shading of the panels (Figure 13a) is a common example of such imbalances. HS issues (Figure 13c) arise when the bypass diodes of shaded cells fail or become isolated. This results in a drop in current and a negative voltage, causing the shaded cells to consume energy from the non-shaded cells instead of generating it. If this condition persists, it can damage the affected solar cells [133] (Figure 13d). Detection methods for HS faults are discussed in [36], with many quick detection techniques relying on infrared measurements to identify the problematic areas.

6.2. Diode Faults

Bypass and blocking diodes are vital components in PV systems, ensuring their efficient and safe operation. Bypass diodes (BpD) help protect the system from reverse voltage while blocking diodes (BkD) prevent reverse current. Common issues with these diodes include short circuits and open circuits, which can occur if a PV string is partially shaded for long periods [134]. The bypass diode is critical for the system’s safe operation [93], but the blocking diode, which is connected in series with the PV module, can sometimes interfere with the proper functioning of Overcurrent Protection Devices (OCPD) [136]. If the reverse current drops below a certain level, known as the Negative Loading Point (LLF), the blocking diode will cut off the current, causing the system to fail. To avoid these types of faults, it is important to carefully choose and thoroughly test both bypass and blocking diodes.

6.3. Junction Box Faults

The junction box (JB) is a critical component for the reliability of photovoltaic (PV) panels during their operation in the field [137]. Corrosion over time can cause the resistance in the JB to increase rapidly [138]. This can lead to electrical arcing between connections, resulting in overheating and melting of the junction box. Such incidents can damage the PV modules and the entire string, leading to significant energy production losses for the system owner. To avoid these issues, several actions and recommendations have been proposed. Regular inspections and maintenance can help detect early signs of wear and corrosion. Ensuring proper installation and using high-quality materials can also reduce the risk of faults. Implementing these strategies can significantly enhance the reliability and lifespan of the junction box, thereby improving the overall performance of the PV system [137].

6.4. PV Module Faults

PV module faults can happen for a variety of reasons, such as corrosion, aging components, internal current leaks, or manufacturing defects. These issues can cause disconnections or internal short circuits within the modules [28,123,124]. When such faults occur, they can pose serious risks like electric shocks or fires. For example, Figure 14 shows various types of defects in PV modules; Figure 14a illustrates a broken glass panel, Figure 14b–d show common problems like oxidation, delamination and bubbles, respectively. These kinds of defects are often seen in PV modules and can affect both the safety and performance of the system. To prevent these issues from escalating, regular inspections and maintenance are crucial. Identifying and addressing these problems early can help maintain the system’s efficiency and longevity.

6.5. Ground Faults

Ground faults in photovoltaic systems (PVSs) happen when there’s an unintended electrical short circuit between the ground and one or more of the system’s energized wires [141]. These faults are a major safety concern because they can create DC electrical arcs at the fault location, which, if left unaddressed, can lead to fires [140]. Detecting ground faults is especially challenging in ungrounded PV systems, as these faults don’t produce enough fault current to be easily detected during normal operations [142]. Ground faults are the most common type of fault in PVSs and can arise from various causes. These include accidental short circuits between wires and the ground, insulation failure of cables and internal ground faults within PV modules (PVMs). A common outcome of ground faults is Potential Induced Degradation (PID), which happens when there’s a significant voltage difference between the cells and the ground. These faults not only pose a fire risk but also affect the overall performance of the system. To mitigate these risks, it’s recommended to use Ground Fault Detectors to monitor insulation in both residential and large-scale PV systems, enhancing safety and reliability [143]. Regular maintenance and timely detection are key to preventing ground faults and ensuring the longevity of PV systems.

6.6. Arc Faults

An arc fault happens when electricity jumps across a gap through the air or another insulating material. These faults come in two main types: (1) Series Arc Faults (AFa), when there is a break or gap in a single electrical wire, and (2) Parallel Arc Faults (AFa), between two wires with potential difference [144]. It is essential for every electrical system to have an Arc Fault Detector to identify these faults.

There are two primary methods for detection [145]:

• DC Method: This method involves monitoring the DC in a wire. By adding a small resistance in series with the circuit, the voltage across the resistor can be measured to detect any anomalies.
• AC Method: this method uses the AC flowing through a wire, with a current transformer acting as a sensor to detect changes caused by an arc fault.

To prevent fires and protect the photovoltaic system (PVS), these detectors need to activate safety circuits when an arc fault is detected. Arc faults pose a significant fire hazard, so detecting and preventing them is critical for the safety and reliability of PV systems. By implementing these detection methods, we can ensure that the system runs safely and efficiently.

6.7. Line-to-Line Faults

Line-to-line faults (LLFs) happen when there is an unintended low-resistance connection between two points with potential different in an electrical system. In PV system, this usually means a short circuit between the cables of different PVMs or arrays with different voltages [146]. These faults in PV arrays can be caused by:

• Cable Insulation Failure: when the insulation around cables deteriorates or fails, it can lead to accidental short circuits between wires.
• Poor Insulation and Mechanical Stress: if the insulation between string connectors is inadequate or if the cables are subjected to mechanical stress, it can result in LLFs.

To prevent these faults, many companies have developed specialized protective devices. These devices detect and mitigate LLFs, ensuring the system remains safe and functional. Along with using these protective devices, regular maintenance and careful installation are key to prevent LLFs and ensure the reliability and safety of photovoltaic systems.

6.8. Relationship of Faults-Performance of PV Systems

The study [149] explores how different environmental factors impact the performance of PV systems. It specifically looks at four factors: (1) dust accumulation, (2) water droplets, (3) bird droppings and (4) partial shading. The key findings from the study are:

• Dust, Shading and Bird Droppings: these factors significantly reduce the current and voltage in PV systems, leading to lower energy production.
• Shading: This has the most significant impact on PV efficiency. When shading covers a quarter, half and three-quarters of the panel surface, the power output drops by 33.7%, 45.1% and 92.6%, respectively.
• Water Droplets: inlike the other factors, water droplets can actually help by cooling the panels, which increases the voltage difference and boosts power output by at least 5.6%.
• Dust: accumulation of dust on panels reduces power output by 8.80% and efficiency by 11.86%.
• Bird Droppings: these decrease system performance by about 7.4%.

These findings emphasize the need for regular cleaning and maintenance of PV panels to ensure they perform at their best. By keeping the panels clean, we can minimize the negative effects of environmental factors and maintain optimal energy production.

In general, environmental factors such as extreme weather conditions, can also severely impact the effectiveness of PV fault detection technologies. Extreme heat can affect thermal cameras to misinterpret overheated areas, while extreme cold may reduce the operational time of drones and cameras by affecting their battery. Rain and humidity can cause obstructions and blur to camera lenses, posing obstacles for accurate faults detection. Strong wind conditions can cause blurry images due to stabilization issues of cameras, while drones could not be able to operate in case of strong winds. Snow and ice may cover the panels, obstructing detection or ever damaging drones and cameras. Finally, direct sunlight may also affect the quality of images due to sun glaring.

7. Evaluation of CV-Based PV Fault Detection Methodologies

Having outlined the fundamental detection and imaging technologies concerning key faults in photovoltaic systems, we now aim to evaluate the effectiveness of the detection methodologies identified in the literature, as imposed by RQ4. For this purpose, a comprehensive table,

Table 5. Evaluation summary of PV fault detection methodologies.

Ref.	Technology	Input Data	Panel Detection Methodology	Fault Detection Methodology	Panel Detection Evaluation (%)	Anomalies Detection Evaluation
[109]	Multispectral imaging	15,330 PV cell images without defects 5915 images with defective cells Training 80% Testing 20%	-	Multispectral (MSI) CNN	-	Accuracy: Thick Line: 76.4% Broken gate: 80.4% Scratches: 48.6% Paste Spot: 82.1% Color diff.: 100% Dirty Cells: 87.2% No Anomalies: 98.1%
[105]	IR Thermography	37 images with 1544 PV cells (Images from UAV)	Creation of background temperature map, automatic thresholding to segment panels from background, removal of unwanted background, estimation of PV panel row orientation, panel dimension correction, preparation for panel analysis	Grid Cell Medians: Division of the panel into a 9 × 10 cell grid and calculation of the median temperature from the individual temperatures in each grid cell.	F1-score: 92.8%	Hot Spots, Hot Substring, Hot Panel (overheat) Average F1-score: 93.9%
[114]	RGB Imaging	Original dataset: 45,754 images Training set: 27,537 Validation set: 18,217	-	Detection Model (ImpactNet), Localization Technique (Mask FCNN) to predict power loss and soiling localization, localization enhancement through BiDIAF, soiling type categorization with WebNN	-	Dust, Snow, Bird Poop, Crack Overall Accuracy: 84.5%
[100]	Visible light camera (CCD) & IR Thermography	- (Images from UAV)	Morphological transformation and Canny Edge algorithm	Thermal imaging and CCD video processing, Hot Pixel-based hot spot detection	-	Hot Spots Cracks & Wear, Delamination Connection Faults
[150]	IR Thermography	A series of flights on a test site (Images from UAV)	Template matching	Template matching	Accuracy: 81%	Hot Spots, Bypass Diodes, Mechanically damaged cells, Fault Contact points Mean Accuracy: 85%
[151]	Visible light camera (RGB) & IR Thermography	15 videos manually annotated for local and general thermal anomalies by three thermal cameras resolutions (Images from UAV)	Image preprocessing to remove noise from the image, Canny algorithm to detect PV edges, Line Separation using Hough Transform, Line Segmentation and Processing, Panel Model Application	Local Hot Spot Detection to detect thermal anomalies within the area of each photovoltaic panel, Global Hot Spot Detection, tracking algorithm to identify and follow the same panels across different frames as the UAV flies over the photovoltaic park	Overall Accuracy: 83%	Local hot spot Accuracy: 73% Global hot spot Accuracy: 85%
[152]	IR Thermography	4.3 million IR images of 107,842 pv panels Panel detection: Training 90%, Testing 10% Anomaly detection: Training 70%, Testing 20%, Validation 10% (Images from UAV)	Panel segmentation through Mask R-CNN	ResNet-50 classifier	Overall Accuracy: 90.01%	Accuracy: Healthy panel: 95.35 ± 0.21% Connection interruption–panel: 98.83 ± 0.42% Short circuit: 66.67 ± 47.14% Connection interruption-string: 100 ± 0% Short circuit string: 83.80 ± 0.76% PID panel: 86.69 ± 1.75% Multiple hot cells: 33.33 ± 23.57% Single hot cell: 57.41 ± 6.93% Hot cells: 80.39 ± 0.26% Diode overheating: 90.06 ± 0.55% Hot spost: 7.07 ± 7.04%
[153]	EL Imaging	148 images of PV cells for the U-net Training: 108 (73%) Testing: 30 (20%) Validation: 10 (7%)	-	Encoder VGG-16 to extract features, Semantic Segmentation with U-net to predict the presence and type of defects	-	Recall: Cracks: 84% Offline areas: 69% Faults in the panel’s conductor lines: 53%
[108]	EL Imaging	47 images of PV panels: 7 healthy panel 40 panels with cracks of different lengths	-	Εnhanced Crack Segmentation (eCS)	-	Cracks from 20 mm up to the entire length of the panel AUC: 91.14%
[154]	IR & RGB Imaging	2038 thermal images (LWIR) for hotspot detection: Training: 1426 (70%) Testing: 306 (15%) Validation: 306 (15%) 1500 low-res visible spectrum digital images (VIS-LR): Training: 1050 (70%) Testing: 225 (15%) Validation: 225 (15%) (Images from UAV)	Canny Algorithm to detect edges of PV modules, Line Separation using Hough Transform, Image rotation optimal detection	YOLOv3	Accuracy: 98%	Accuracy: hotspot: 80.30% hotspot on junction box accuracy: 90.27% puddle accuracy: 82.48% bird dropping accuracy: 81.97% raised panel: 84.00% delamination: 93.61% strong soiling: 73.75% soiling accuracy: 90.00%
[155]	RGB Imaging	126 images of multiple defects on PV panels: Training 66.6% Testing 33.3% (Images from UAV)	-	Detecting anomalies with the Kirsh Operator image segmentation technique, the trained CNN extracts feature vectors of anomalies, the resulting anomaly vectors are inserted into a Multi Class-SVM which classifies 5 final anomalies	-	Accuracy: Dust shading: 97.63% Encapsulant delamination: 98.59% Glass breakage: 98.42% Gridline Corrosion: 95.84% Snail trails: 95.03% Yellowing: 97.76%
[156]	EL Imaging	Dataset 19,228 EL images 640 × 512 For YOLO model 1025 images used: Training: 762 (74.5%) Testing: 134 (12.5%) Validation: 134 (13.0%)	Automatic Perspective Transform, Automatic Cell Segmentation to identify cell boundaries, UNet to extract panel features, OpenCV for Line and Corner Detection.	Object Detection with YOLOv3 Model, Image Classification with ResNet18, ResNet50 and ResNet152 models to classify cells into 4 types of anomalies (cracks, intra-cell defects, oxygen induced defects and solder disconnections)	Accuracy: 98.6%	Average F1-score: YOLO: 78% ResNet18: 83%
[157]	EL Imaging	PV Multi-Defect dataset: 305 images 5800 × 3504 of 5 types of anomalies. After preprocessing, 1108 anomaly images: 80% for Training 20% for Testing and Validation	-	Ghost convolution with BottleneckCSP YOLOv5 (GBH-YOLOv5)	-	mAP: Broken Glass: 99.5 ± 0.01 Hot Spot: 97.5 ± 0.02% Black_Border: 97.2 ± 0.02% Scrath: 97.4 ± 0.02% No_Electricity: 98.0 ± 0.02%
[117]	IR Thermography	18 videos, of which: 13 (72%) for Training 5 (28%) for Testing (Images from UAV)	YOLOv2 and YOLOv3: Image Inclusion, Image Division, Bounding Box Predictions	-	YOLOv2: Accuracy 89% YOLOv3: Accuracy 91%	-
[158]	RGB Imaging	3150 images with 6 anomaly classes (Images from UAV)	-	AlexNet for Feature Extraction, J48 decision tree for Feature Selection, Classification with k-nearest neighbors (kNN): Locally weighted learning (LWL) and K-star are compared	-	Accuracy: Delamination: 99.61% Burn marks: 97.90% Discoloration: 98.85% Snail Trail: 99.61% Glass Breakage: 99.61% Good Panel: 98.09%
[159]	EL Imaging	UCF EL Defect Dataset inluding 17,064 EL images: 80–20 ratio for training and testing/validation	-	Semantic Segmentation with DeepLabv3 and ResNet-50 as backbone	-	Accuracy: No Defect: 98% Crack: 81% Contact: 66% Interconnection Interruption: 26% Corrosion: 69%
[160]	EL Imaging	6264 images: 5011 images (80%) for training, 1253 images (20%) for testing	-	Unsupervised ML–Principal Component Analysis—PCA to reduce the dimensionality of image data, Hierarchical Clustering to group images based on features similarity, Feature Extraction–Haralick Feature, Supervised ML–CNN and SVM classification	-	Defects: Cracks, Busbar corrosion, Dark spots, Clear or in good condition Mean accuracy of Models: SVM: 98.95% CNN: 98.24%
[161]	IR Thermography	Infrared Solar Modules dataset: 20,000 IR images: 10,000 with no anomalies 10,000 with 11 categories of anomalies: For feature extraction the model Efficientb0 used was pre-trained. (Images from UAV) For classification in SVM, 80% was used for Training/20% for Testing	-	The Efficientb0 model for feature extraction, Network Component Analysis (NCA) method to select most significant features, Classification with SVM classifier	-	F1-scores: Hot-Spot: 88.05% Multiple Cells Hot-spot: 84.27% Cracks: 91.40% Active bypass diode: 97.51% Diodes: 95.04% Thin film hot-spot: 84.45% Multiple film hot-spots: 85.89% Offline module: 90.93% Shadowing: 91.01% Soiling: 82.17% Vegetation: 89.30% No anomaly: 97.85%
[162]	EL Imaging	3629 images, 2129 defective and 1500 non-defective: Training: 847 defective images and 452 non-defective images	-	Bidirectional Attention Feature Pyramid Network (BAFPN), Multi-head Cosine Non-local Attention Module, Embedding of BAFPN into Region Proposal Network (RPN) in Faster RCNN+FPN	-	Classification: F-score: 98.70% Detection: mAP: 88.7%
[163]	IR Thermography and RGB Imaging	240 panel images: 80% for Training, 20% for Testing	Region proposal by Maximally Stable Extremal Regions (MSER) + filtering by size	Segmentation by binary thresholding	-	Accuracy: Hot spot: 97%
[164]	IR Thermography	1171 panel images with hot spots (Images from UAV)	Edge extraction by Hough transform + postprocessing	Segmentation by binary thresholding	F-score: 69%	Hot Spot F-score: 59.0%
[23]	IR Thermography and RGB Imaging	34 visual and 34 IR images (Images from UAV)	From visual images, module recognition, mosaicking, numbering and counting	From IR images, image filtering and elaboration, defect identification	-	-
[165]	IR Thermography	panel images with one anomaly class (Images from UAV)	Template matching	Template matching	F-score: 83.0%	Hot Spot F-score: 75.0%
[166]	IR Thermography	100 thermal images: Training 80%, Testing 20% (Images from UAV)	Rectangle extraction by adaptive thresholding + SVM classifier on texture features	-	F-score: 98.9%	-
[167]	IR Thermography	798 panel images, with 398 images of 4 class anomalies and 400 non-defective images: Training 80%, Testing 20%	-	Defect classification: SIFT feature extraction + RF classifier, VGG16 and MobileNet	-	Accuracy: Feature-based: up to 91.2% DL models: up to 89.5%
[168]	IR Thermography	235 panel images: Training 92%, Testing 8% (Images from UAV)	DL semantic segmentation (ResNet-34+U-Net)	-	F-score: 97.11%	-
[169]	IR Thermography	Dataset of frames of videos recorded in grayscale (Images from UAV)	-	Segmentation by VGG-16 based DL model	-	Hot spot, disconnections (strings and substrings)
[170]	IR Thermography	3336 thermal images, with 811 of damaged and 2525 of normal PV cells: 80% Training, 20% Testing (Images from UAV)	-	DCNN (training by VGG-16) of entire video frame	-	2 classes: defective (e.g., hot spot), normal mean F1-score: up to 69.0%
[171]	Near-infrared EL image	PVEL-AD-2021 benchmark dataset	-	Partial Convolution and Switchable Atrous Convolution YOLOv7	-	Precision: 88.3%
[172]	RGB imaging	Solar panel soiling image dataset of 45,469 images	-	Vision transformer (ViT)	-	Accuracy: 97%
[173]	EL imaging	Training with 2018 images of bright and 101,376 of non-bright hot spots patches	-	Feature extraction and generative adversarial networks (GANs)	-	F1-score: 93%
[174]	RGB imaging	4500 PV defect datasets including cracks, broken grids, black cores, thick lines and hot spot	-	Faster-RCNN and YOLOv5	-	mAP: Faster-RCNN: 92.6% YOLOv5q 91.4%
[175]	EL imaging	PVEL-AD dataset	-	YOLOv4 with an improved Convolutional Block Attention Module (YOLO-iCBAM)	-	F1-score: 71.6% mAP: 74.8%
[176]	EL imaging	593 cell images, 80,000 images	-	C2f module in YOLOv8 to replace the C3 module in the backbone network	-	mAP: 67.5%
[177]	IR Thermography	Thermal camera mounted on a UAV (Images from UAV)	-	Image processing: contour defining, color/pixel selection	-	Accuracy: 75%
[178]	RGB imaging	2624 grayscale images of solar cells of two classes	-	Decision Tree, SVM, KNN, Ensemble and Discriminant	-	Accuracy: up to 98.34% with Ensemble
[179]	EL imaging	Global public dataset of EL images of Hebei and Beijing University (80–20 split)	-	YOLOv8	-	Average precision: 90.5%
[180]	IR Thermography	Database from a solar power plant of 42,048 modules (Images from UAV)	-	Mask R-CNN	-	mAP: 72.1%
[181]	EL imaging	584 of normal (300 × 300) and 197 images of abnormal solar cells		GAN and auto-encoder (AE)		Accuracy: 90%

, has been structured, linking all the primary automated detection technologies with the use of computational intelligence algorithms, for detecting each specific set of anomalies in PV systems in the literature. The table records the performance of each methodology achieved in each study, in relation to (1) the ability to identify photovoltaic panels and (2), the performance of each methodology in detecting a corresponding set of anomalies.

However, since Table 5 is extensive and conclusions cannot easily draw, a color-coded evaluation table (

Table 6. Evaluation summary of PV fault detection methodologies (Performance color code: red for performances 0–30%, yellow for 30–70%, green for 70–100%, gray for no numerical performance reported). Shading anomalies are marked in blue font.

Ref.	Hot spots	Cracks	Glass Breakage	Scratches	Delamination	Discoloration	Corrosion	Connection Faults	Short Circuit	PID Ground Faults	Diode Faults	Burn Marks	Inactive Cells	Dust	Snail Trails	Bird Poop	Snow	Pamel Detection	UAV Inspection	Rank
[109]		X	X		X		X		X	X		X	X		X	X	X	X	X	5
[105]		X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X		✓
[114]	X		X	X	X	X	X	X	X	X	X	X	X		X			X	X
[100]			X	X		X	X		X	X	X	X	X	X	X	X	X	X	✓
[150]		X	X	X	X	X	X	X	X	X		X		X	X	X	X		✓
[151]		X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X		✓
[152]		X	X	X	X	X	X					X		X	X	X	X		✓	1
[153]	X		X	X	X	X	X		X	X	X	X		X	X	X	X	X	X
[108]	X		X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X
[154]		X	X	X		X	X		X	X	X	X	X		X		X		✓	3
[155]	X	X		X				X	X	X	X	X	X			X	X	X	✓	2
[156]	X		X	X	X	X		X	X	X	X	X		X	X	X	X		X
[157]		X			X	X	X		X	X	X	X		X	X	X	X	X	X
[117]	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X		✓
[158]	X	X		X			X	X	X	X	X		X	X		X	X	X	✓	4
[159]	X		X	X	X	X			X	X	X	X		X	X	X	X	X	X
[160]	X		X	X	X	X		X	X	X	X		X	X	X	X	X	X	X
[161]			X	X	X	X	X	X	X	X		X		X	X	X	X	X	✓
[162]			X	X		X	X	X	X	X	X	X	X	X	X	X	X	X	X
[163]		X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X
[164]		X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X		✓
[23]			X	X	X	X		X	X	X	X	X	X	X		X	X		✓
[165]		X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X		✓
[166]	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X		✓
[168]	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X		✓
[169]		X	X	X	X	X	X		X	X	X	X	X	X	X	X	X	X	✓
[170]		X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	✓
[171]			X	X		X	X	X	X	X	X	X	X	X	X	X	X	X	X
[172]		X	X	X	X	X	X	X	X	X	X	X	X		X			X	X
[173]		X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X
[174]				X		X	X	X	X	X	X		X	X	X	X	X	X	X
[175]			X	X		X	X	X	X	X	X	X	X	X	X	X	X	X	X
[176]			X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X
[177]	X		X	X	X	X	X			X		X	X		X	X	X	X	✓
[178]	X		X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X	X
[179]	X		X	X			X	X		X	X		X	X	X	X	X	X	X
[180]		X	X	X	X	X	X			X	X	X		X	X	X	X	X	✓
[181]			X	X	X	X	X		X	X	X	X	X	X	X	X	X	X	X

) was also created. More specifically, the information from the cumulative Table 5 which connects the performance of each identified technique with the corresponding faults, has been coded by using four color levels: red for performances 0–30%, yellow for 30–70%, green for 70–100%, gray when no numerical performance is reported. UAV-based applications were also considered. Next, three evaluation criteria were selected:

• The total number of green indicators, showing the range of faults that the methodology effectively addressed (degree of satisfactory for anomaly coverage).
• Whether it satisfactorily detects shading anomalies (dust, snail trails, bird droppings and snow deposits), which significantly affect cell performance, as already discussed.
• The degree of automation provided by the methodology, consisting of two sub-questions:

  a.

  Does the methodology in this study detected photovoltaic units? (Column: “Panel Detection”)

  b.

  Was the automated image acquisition in this study via UAV technology? (Column: “UAV Inspection”)

By applying these criteria to each row of Table 6, we rank the top five optimal methods from our literature findings in the final column. The latter approach aims to identify and evaluate the best performing approaches of the literature according to preset criteria.

It should be noted that cost-effectiveness and economic viability of each proposed PV fault detection method is crucial, especially towards their large-scale adaption in the solar industry. Yet, the comparison of the various methods could not include such cost information since the latter was not reported within the articles and therefore was not available. Moreover, the cost-effectiveness of camera-based PV panel inspection systems is not easily comparable since the viability of each method depends on several factors, such as the type of the used cameras, and the size of the solar installation that will be applied to. Low cost RGB cameras can be used, however hotspots and other defects can be better detected by thermal cameras which are most cost effective. For cameras mounted on drones, high resolution is essential, which are more pricey. Combined systems integrating both thermal and high-resolution cameras are more advanced and expensive options that may offer enhanced detection capabilities. Yet costs are closely related to the size of the solar installation. Ground based inspection systems can be more affordable, yet they are not feasible solutions for large scale PV farms. Drone-based systems on the contrary, can be affordable for large scale PV installations regardless costly mounted high-resolution cameras and thermal imaging equipment, while the latter would not be economically beneficial to use it in small PV installations. Therefore, in order to compare and choose the most cost-effective method, all the above aspects should be carefully considered according to each application case.

8. Discussion

In this holistic literature review, we analyzed existing methodologies for fault detection in photovoltaic systems using computer vision, machine learning and deep learning techniques. The main observations and conclusions of the study are summarized as follows:

• Evolution of Detection Technologies: The advancement of computer vision algorithms in recent years has significantly improved the accuracy and effectiveness of fault detection. In particular, DL techniques, especially CNNs, have proven to be highly efficient in detecting anomalies in photovoltaic units.
• Shading Anomalies: Fault detection and regular maintenance of photovoltaic panels are critical for maintaining optimal performance. Environmental factors that increase panel shading, such as various types of dust and bird droppings, can significantly impact system performance, making continuous monitoring and cleaning essential.
• Methodology Evaluation: From the evaluation of various methodologies, the CNN variant Resnet 50 shows a very promising future among the other literature findings. It is also evident that DL techniques, such as YOLO variants and combinations of CNN with SVM classifiers, provide high detection accuracy and cover a wide range of faults. However, traditional methods, such as thresholding techniques, while satisfactory in some cases, have limitations in covering different types of anomalies.

To further improve fault detection in photovoltaic systems and develop more efficient and economically viable solutions, the following directions are proposed:

• Scalability and Economic Viability: The development of scalable deep learning models that can incorporate economic analyses will be important. This will enable the application of these technologies on a large scale and ensure the viability of solutions in large photovoltaic parks.
• Real-Time Detection: Adapting and improving real-time detection techniques is crucial. Developing algorithms that can operate effectively under various environmental conditions and provide immediate feedback will enhance maintenance efficiency and the performance of photovoltaic systems.
• Integration of Multiple Technologies: Integrating technologies such as infrared imaging, electroluminescence imaging and RGB imaging, combined with deep learning algorithms, will allow for better fault detection and diagnosis. Hybrid models that could combine different AI techniques could also be investigated to improve the accuracy of PV fault detection, as well as the integration of IoT and edge computing devices for continuous real-time data collection. All the above could lead to more comprehensive and reliable detection solutions.
• Data augmentation and synthetic data generation: data collection can be also enhanced through augmentation techniques and the generation of synthetic data, towards creating balanced benchmark datasets of all kinds of faults to effectively train AI models.

With ongoing research and development in these areas, fault detection in photovoltaic systems is expected to improve significantly, offering higher performance and longer lifespan for photovoltaic systems. Therefore, practical implications of vision-based PV systems that could be translated into actionable insights for PV industry stakeholders, include the following:

• Cost management due to the early detection of faults, preventing minor faults to expand and ruin the entire system, therefore reducing repairing and replacement costs for PV farms. Data analysis could also aim towards preventing faults before occurring through the identification of patterns, resulting in PV installations performance optimization. In general, data-based insights can overall improve predictive maintenance strategies of PV installations, due to the ability to prevent, plan and act.
• Efficiency of operations for large PV installations, since automated monitoring solutions can reduce inspection time, as well as both human labor and human error.
• By reducing human involvement, safety is provided for PV installation personnel, since inspections in the field involve hazards due to the nature of the PV structures that require the worker to climb ladders or high places in order to properly inspect, as well as their exposure to potentially adverse environmental conditions.
• Safety is also a requirement for PV system operations. Fault detection aims towards compliance with regulations, i.e., standards and requirements for PV installations, to ensure their operation safety as predefined by corresponding guidelines.
• Computer-vision PV fault detection strategies may lead to more consistent, reliable and efficient PV inspection, making PV installations that adapt such technologies more competitive in the market.

More analytically, regarding the color-coded evaluation of Table 6, we observe that the detection methodology from the study [152] ranks first. This methodology was based on infrared thermography with UAVs, using Mask FCNNs for detecting photovoltaic units, and employed the DL network Resnet 50 to detect and classify seven out of the total seventeen anomalies. Of these, six were detected with satisfactory accuracy and one with average (less than 70%) accuracy.

The second highest performance according to our criteria, goes to the methodology presented in [155], which utilizes RGB imaging via UAVs, CNN for feature extraction and SVM for anomaly classification. This method also satisfactorily covers six out of seventeen faults. However, it lacks an integrated algorithm for detecting photovoltaic units, which reduces its level of automation compared to the top-ranked method. It is worth noting that of the six faults detected with satisfactory accuracy, two are shading anomalies (dust accumulation and snail trails). Given our observations on our second research finding—shading anomalies significantly impact the performance and output of PV units—this could potentially place the methodology in the first rank depending on the priorities of each specific implementation (e.g., resource allocation relative to desired PV unit performance). Accordingly, we could continue to rank all other methodologies based on the overall degree of satisfactory anomaly coverage, anomaly detection capability and the level of automation included in each specific study.

Another significant observation from the evaluation results is that the top five methodologies all involve DL techniques, with the top three including CNNs or their variations and two using YOLO variations. In contrast, segmentation methodologies with thresholding, while providing satisfactory accuracy, are able to detect a narrow range of different anomaly types. It is also noteworthy that the SVM classifier was found in the second position. This shows that even when combined with a CNN for feature extraction, SVM remains a highly competitive ML method for effective anomaly detection in photovoltaic units, despite the current trend towards using DL algorithms in computer vision tasks.

Practical applications and real-world implementations of vision-based PV fault detection can be found solely for large scale PV installations. The most well-known PV power station is that of Kamuthi in Tamil Nadu region in India. The latter is the largest PV farm, spreading across an area of 10 km². The farm employes drones equipped with high-resolution cameras and thermal sensors to overall monitor the status of all PV modules. Drones fly over the PV farm to collect images, and data analysis is performed afterwards to detect anomalies and various types of faults. Challenges in real-world implementations include as already mentioned in Section 6.8, mainly the effects of environmental conditions that influence the accuracy of cameras, and the demanding processing of the vast number of generated data

9. Conclusions

In this work, a comprehensive review was conducted focusing on vision-based monitoring techniques for fault detection of PV systems. The research was guided by four research questions, aiming to identify (1) the evolution of computer vision algorithms for PV fault detection within the decade, (2) to present all key monitoring techniques for PV fault detection, along with an individual and comparative assessment of their capabilities and limitations, focusing on AI-based machine vision techniques, (3) to identify the common faults in PV system and how they affect the systems’ performance and finally (4) to conclude to the most efficient CV-based methodologies through the review of the literature. Results indicated that among seventeen identified faults, shading anomalies are those that significantly impact the performance of PV units, while the top five CV-based PV fault detection methodologies, according to preset evaluation criteria, involve deep learning methods, such as CNNs and YOLO variations.

Key findings reveal that advancements of computer vision algorithms in recent years have significantly improved the accuracy of PV fault detection, indicating CNNs as highly efficient performing models for PV fault detection. Moreover, DL techniques, such as YOLO variants and combinations of CNN with SVM classifiers, can provide high detection accuracies and cover a wide range of different faults. Research concluded that fault detection and regular maintenance of photovoltaic panels are critical for maintaining optimal performance, while environmental factors are those to mainly affect the systems’ performance.

It is evident from the literature that research on PV fault detection is ongoing, and it is expected that future technological developments, such as powerful processors to run AI-powered CV algorithms in real-time on UAVs, will further enhance this area of research, offering higher efficiency and longer lifetime for PV systems. Future research direction stemming from this work suggest the development of cutting-edge CV algorithms that could combine different sensing modalities, such as thermography and electroluminescence, towards more efficient PV fault detection. Hybrid models that could combine different AI techniques could also be investigated to improve the accuracy of PV fault detection. Real-time monitoring is a requirement in fault detection; therefore, efforts could focus on the integration of IoT and edge computing devices for continuous real-time data collection. Data collection can be also enhanced through the generation of synthetic data, towards creating balanced benchmark datasets of all kinds of faults to train AI models.