Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
Scientific Reports volume 16, Article number: 10056 (2026) Cite this article
1482 Accesses
Metrics details
This research introduces a novel artificial intelligence (AI) framework for fault detection and diagnosis (FDD) in photovoltaic (PV) systems that combines Convolutional Neural Networks (CNNs) with time–frequency analysis via the Wigner–Ville Distribution (WVD). The proposed method transforms raw numerical measurements—including solar irradiance, temperature, voltage, current, and power—into compact 6 × 12 time–frequency image representations, enabling effective spatial feature extraction by CNNs that are well suited to image-like data. The framework is benchmarked under both noiseless and noisy conditions on a comprehensive 17‑class dataset comprising one healthy condition (C0) and sixteen fault types (F1–F16), including progressive short‑circuit faults within a single string, pure partial‑shading faults, and combined inter‑string short‑circuit and asymmetric partial‑shading patterns along PV strings. To contextualize performance, the CNN–WVD model is compared not only with classical Artificial Neural Networks (ANNs) and Deep Neural Networks (DNNs) but also with Gradient Boosting Machines (GBM), Random Forests (RF), Support Vector Machines (SVM), and k‑Nearest Neighbors (kNN), all trained on the same WVD‑transformed data. In noiseless conditions, ANN and DNN achieve 99.51% and 99.49% accuracy, respectively, while the CNN attains 97.09%; RF, SVM, GBM, and kNN reach 93.47%, 88.62%, 84.01%, and 75.69% accuracy. Under noisy conditions that emulate real PV environments, the CNN is the most robust model with 90.27% accuracy, outperforming ANN (82.20%), RF (82.80%), SVM (83.85%), GBM (73.85%), DNN (76.27%), and kNN (72.80%). Key contributions include: (i) the use of WVD to obtain highly informative time–frequency representations of PV electrical signals, (ii) a structured data‑organization strategy that maps multivariate PV measurements into fixed‑size WVD images, and (iii) a CNN architecture that preserves high discrimination capability across closely related fault severities and locations, even in the presence of noise achieving 90.8% accuracy under realistic sensor noise ((1 times) baseline uncertainty: (pm 10{W mathord{left/ {vphantom {W {m^{2} }}} right. kern-0pt} {m^{2} }}) irradiance, (pm 2 , C) temperature, (pm 5 , V) voltage, (pm 1 , A) current, (pm 25 , W) power) and maintaining 71.5% accuracy at (3 times) noise, representing extreme aging sensor conditions. With a competitive degradation of only 8.91 percentage points—lower than the neural-network baselines (ANN: 16.27%, DNN: 15.00%) and the tree ensemble RF (11.34%)—the CNN + WVD framework demonstrates superior noise robustness for long-term deployment in real-world PV installations. By bridging advanced time–frequency analysis with deep learning and systematically comparing against a broad set of machine‑learning baselines, the proposed framework enables fully automated, fine‑grained PV fault classification without manual feature engineering, thereby enhancing monitoring reliability, reducing downtime, and supporting predictive maintenance in large‑scale PV deployments.
Photovoltaic (PV) systems have seen rapid global growth, but faults in these systems pose a major threat to their reliability and efficiency. Visual inspection methods using techniques like infrared thermography, electroluminescence, and photoluminescence have been widely used to detect PV faults. Recently, deep learning (DL) approaches, particularly convolutional neural networks, have shown promising results in automating visual PV fault detection.
A DNN-based failures detection and diagnosis approach to identify bird droppings on PV modules is described in¹. The network is trained using aerial imagery and segments pixels using an encoder-decoder architecture. A proposed method presents an identification of the imperfections within a local coordinate system. The degree of discolouration and the impact it has on the performance of the solar module are taken into consideration when ranking the damage that occurs during post-processing. For the purpose of determining how effective the system is, the influence that bird droppings have on the parameters of the modules is being investigated.
The identification of flaws within a local coordinate system is accomplished through the utilization of a positioning method. During the post-processing stage, the severity of the discoloration and the impact it has on the output of the PV module are taken into consideration when ranking the degrees of damage. An analysis of the effects that bird droppings have on the parameters of the module is used to determine whether the system is effective.
A novel Fault Detection and Diagnosis (FDD) model has been developed through research, which integrates the architectures of Convolutional Neural Networks (CNN) and Wasserstein Generative Adversarial Networks (WGAN). Three interconnected modules comprise the model: discriminator, generator, and classifier. The analysis of PV data is conducted using a two-dimensional framework, which also produces supplementary labelled samples to augment the training dataset for the CNN-based classifier. The model exhibited precise capability in identifying and diagnosing line-line and open circuit faults through experiments conducted on a laboratory grid-connected PV system. By making this research a contribution to the ongoing discourse on defect diagnosis in PV systems, additional avenues for investigation are opened².
An economical method for PV plant real-time monitoring and diagnosis is proposed in³. Partial shading conditions (PSC) are detected utilising real-time scalograms from a PV data collecting device by a pre-trained AlexNet CNN. This technique is implemented on a PyBadge microcontroller, the model achieves a 98.05% fault detection accuracy. Under experimental verifications demonstrate that the system can remove unwanted objects, reduce thermal problems, and automatically disconnect partially shaded panels, guaranteeing real-time restoration of normal PV system operation.
By combining manual and automatic feature extraction methods, a study developed a deep learning model to improve the accuracy of PV image classification⁴. This model uses a DNN, and histogram of directional gradient (HoG) features that extracted from PV images. The classification accuracy of this model compared with six other existed methods. Applying this methodology not only improves PVS monitoring and maintenance, but also enhances the long-term feasibility of renewable energy technologies. The new methodology of this study lays the foundation for subsequent advances in PVS monitoring and maintenance, ensuring a reliable and effective technique for evaluating PV system performance.
In the field of PVS, the Squeeze-and-Excitation (SE)-inception Multi-Input CNN⁵ is a novel approach that has been proposed for the purpose of detecting Series Arc Failures (SAFs). Denoising techniques are utilized in this approach to decrease the influence that switching frequencies have on the predictive accuracy of diagnostics. The SE network topology is integrated with the channel attention mechanism to accomplish this goal. The fact that this technique is very resistant to disturbances and can detect SAFs in a wide range of situations makes it a significant advancement in the field of PVS detection. A defect diagnosis scheme for DC side PV arrays (PVA) is presented in⁶ by employing a Dual-Channel CNN (DcCNN). This algorithm utilizes the current and voltage electrical time series graph to extract features in order to enhance the model’s ability to diagnose PVA defects under different conditions, where a Feature Selection Structure (FSS) is implemented. The FSS not only mechanically extracts significant features but also assesses them in preparation for subsequent classification. A thorough experiment was conducted on a laboratory roof grid connected PVS to illustrate how the proposed approach outperformed.
For detecting hot areas on an Internet of Things (IoT) platform, a novel method is suggested that employs image analysis with two consecutive CNNs⁷. By integrating thermal cameras into unmanned aerial vehicles, the proposed method identifies thermal patterns that are indicative of malfunctions in photovoltaic panels. With a decrease in false positives, the platform processes data automatically via various CNNs, accomplishing 99% panel detection and 96% hot spot detection. Dust is accounted for in a novel fault-diagnosis method for PVA⁸. Resampled PVA characteristic curves are transformed and normalized using the Isc-Voc normalized Gramian-Angular-Difference-Field method, which also generates transformed graphical feature matrices. In order to identify intricate fault types, a CNN model is implemented alongside convolutional block attention modules for the purpose of fault classification. The approach exhibits exceptional precision and dependability in defect diagnosis across a wide range of operational circumstances, thereby surmounting the constraints of existing methodologies. A 3D CNN for PV fault detection and classification is described in⁹. The Gramian Angular Field transform is utilized to convert direct current and alternating current signals to three-dimensional images. In terms of overall accuracy and classification, this method surpasses alternative machine learning approaches, including k-nearest neighbor, Random Forest, Decision Tree, and Support Vector Machine. The rapid growth of decentralized photovoltaic systems presents monitoring and maintenance challenges. The authors in¹⁰ investigates the application of CNN and VGG-16 models for defect diagnosis with thermographic images. The research discovered that the VGG-16 model, which had been fine-tuned, achieved an average fault detection accuracy of 99.91% and a fault diagnosis accuracy of 99.80% for five defects. The experimental evaluations demonstrated that the fine-tuned model achieved a high level of accuracy, whereas the small-DCNN model exhibited a slightly lower level of accuracy.
Recent advances in time–frequency analysis for fault diagnosis further support the methodological choices adopted in this work. In¹¹ introduced an enhanced analog–circuit diagnosis scheme based on continuous wavelet transform and dual–stream convolutional fusion, showing that multi–scale time–frequency representations substantially improve feature separability for subtle fault signatures., Our WVD–based representation follows the same principle of joint time–frequency analysis, but provides higher energy concentration and resolution, avoiding the scale–resolution trade–off that is intrinsic to wavelet kernels.​
Complementary progress in energy forecasting also underlines the value of frequency–domain information. An OOA–optimized bidirectional LSTM with spectral attention for short–term power load forecasting is proposed in¹², where spectral features significantly boosted prediction accuracy., In parallel, in¹³demonstrated that modern machine–learning pipelines can reliably model power consumption patterns in complex energy systems, highlighting the central role of data–driven methods in operational decision making. Wind and power–quality studies provide additional evidence for combining domain–specific signal processing with lightweight neural networks. The NRBO–TXAD framework for wind–speed forecasting¹⁴ illustrates how meta–heuristic optimization can enhance neural architectures in renewable–energy applications,, while Wang et al.¹⁵ recently introduced physics–inspired time–frequency feature extraction with compact neural networks for power–quality disturbance classification, achieving high accuracy with low computational cost., This latter result is particularly aligned with our strategy of pairing interpretable time–frequency descriptors with efficient deep models for PV fault diagnosis.​ Taken together, these recent contributions converge on a consistent message: high–resolution time–frequency representations, when coupled with tailored neural architectures, provide a powerful and increasingly accepted foundation for robust diagnostics and forecasting in modern energy systems. Our WVD–driven PV fault–detection framework fits naturally within this emerging research direction and extends it to safety–critical photovoltaic operation.
The reviewed research show both advantages and drawbacks of using deep learning methods for diagnosing and detecting PV faults: Among these advantages are the great accuracy attained by techniques such as Generative Adversarial Networks and CNNs, as well as the combination of location and severity analysis to accurately identify and measure flaws. Additionally, some solutions include real-time and cost-effective monitoring features. However, the requirement for large amounts of training data, the possibility of false positives, the reliance on specialized technology, and the absence of knowledge regarding real-world implementation and ongoing maintenance are still problems, though. To overcome these obstacles and enable the PV industry’s broad adoption of these defect detection and diagnosis technologies, more research is required. Concurrent fault detection and identifying multiple faults occurring simultaneously in a PV system—is a significant challenge due to the complexity of PV systems and wide range of potential issues. Conventional methods typically detect only single faults at a time. Deep learning approaches like CNNs have shown promise for concurrent fault detection. CNNs can be trained on datasets of PV system images with pre-existing faults to classify new unseen images and identify the presence of multiple anomalies. Other deep learning models used for concurrent fault detection include RNNs, autoencoders, GANs, SVMs, and decision trees. These integrate features from multiple sensor data sources to improve fault detection accuracy. Key challenges include the need for large labeled image datasets, efficient training techniques, and developing robust models that can handle noisy inputs. Integrating data from various sensors can help manage uncertainty and improve overall fault detection performance.
In this work, the implementation methodology of the proposed approach involves several sequential steps. Initially, numerical data encompassing parameters such as solar irradiance, temperature, voltage, current, and power at maximum point undergoes a transformation facilitated by the Wigner-Ville distribution. This transformation converts the numerical data into image representations, thereby rendering it suitable for subsequent processing by CNNs. Following this transformation, the data is structured into 6 × 12 matrices, interpreted as images, where each row corresponds to a distinct parameter and each column represents specific time intervals or data points. This organized arrangement facilitates the subsequent application of CNNs for fault detection and diagnosis. Subsequently, the CNN model undergoes training utilizing labeled data to discern patterns associated with different fault conditions. During the training process, the model changes its parameters over and over to reduce the difference between what it thinks is a fault and what it actually is. This makes it better at finding and diagnosing faults in PV systems. Post-training, the trained CNN model undergoes evaluation and validation using separate datasets to assess its performance in fault detection and diagnosis. This evaluation involves subjecting the model to unseen data to determine its accuracy, sensitivity, specificity, and other pertinent performance metrics. Additionally, the robustness of the method to varying noise levels is scrutinized to ascertain its reliability in real-world PV system environments. Following this thorough method, the suggested method skillfully uses the Wigner-Ville distribution and CNNs to improve fault finding and diagnosis in PV systems, providing a promising way to get around the problems that come with traditional methods.
The proposed approach presents a promising method for improved fault detection and diagnosis in PV systems compared to traditional techniques. The key advantages of the proposed approach are:
Transformation of numerical PV data into image representations using the Wigner-Ville distribution. This allows the use of powerful CNN models for fault analysis.
Structured organization of the data into 6 × 12 matrices, with each row representing a parameter and each column a time interval. This facilitates effective CNN processing.
CNN model training on labeled fault data to learn patterns associated with different fault conditions, improving fault detection and diagnosis capabilities.
Rigorous evaluation of the trained CNN model on unseen data to assess its accuracy, sensitivity, specificity, and robustness to noise. This ensures reliable performance in real-world PV environments.
Compared to traditional fault detection methods, the proposed approach leveraging the Wigner-Ville transform and CNNs provides several advantages:
Automated fault detection and diagnosis without manual interpretation
Improved accuracy and reliability in identifying complex, overlapping fault conditions
Robustness to noise and environmental variations in PV systems
Potential for early fault prediction and preventive maintenance
Overall, the methodical implementation and thorough evaluation demonstrate the promise of this CNN-based approach to overcome limitations of previous PV fault detection techniques, enabling more effective monitoring and maintenance of PV systems.
The prototype PV system under investigation consists of essential components for comprehensive monitoring and fault detection capabilities¹⁶. Figure 1 illustrates the experimental setup of the photovoltaic study plant and its comprehensive monitoring system. At the heart of the system is a grid-connected PV array comprising 30 panels, with 15 panels connected in series. To capture critical performance parameters, the setup includes a PV inverter (DC/AC), an in-plane and horizontal irradiance sensor, and a K-type thermocouple temperature sensor. These sensors feed data to a data acquisition system, specifically an Agilent 34970A¹⁷. The acquired data is then analyzed on a computer to enable the detection and diagnosis of any faults or issues within the PV array¹⁸. By closely monitoring parameters such as irradiance, temperature, voltage, and current, this work aims to develop and test a robust methodology for identifying and understanding the root causes of potential faults in this PV system prototype. The comprehensive data collection and analysis approach is key for enhancing the reliability and performance of grid-connected PV installations.
The PV study plant and the monitored system.
The PV cell, the fundamental building block of solar energy conversion, can be accurately represented by an equivalent circuit model that captures its key electrical characteristics. Figure 2 presents the equivalent electrical circuit of the ODM used to characterize the photovoltaic cell behavior. This model, widely adopted in the literature, consists of several essential components. The photocurrent (({I}_{ph})) represents the current generated by the cell when exposed to sunlight, a parameter that determines the output power. The diode (({I}_{d})) models the nonlinear behavior of the PV cell, mimicking the action of a semiconductor junction. In parallel, the shunt resistance (({R}_{sh})) accounts for the imperfections within the cell structure, providing an alternative current path. Lastly, the series resistance (({R}_{s})) models the resistance to the flow of current through the solar cell materials. By considering these elements; photocurrent, diode, shunt resistance, and series resistance, the equivalent circuit can accurately predict the terminal power (({P}_{pv})) and the output current ((I_{pv})) and voltage ((V_{pv})) produced by the PV cell under varying operating conditions, such as changes in light intensity and temperature. This comprehensive equivalent circuit model serves as a powerful tool for analyzing, simulating, and optimizing the performance of photovoltaic systems.
The one diode model of PV cell.
According to the previous Fig. 2 and based on the current divided the PV currents is defined by the following equation:
where:
(q) is a constant that denotes the absolute value of an electron’s electric charge (1.6 × 10^–19 C);
(k_{B}) is the Boltzmann constant (1.3 × 10^–23 J/K);
(T_{p}) represents the ambient temperature.
The parameters should be identified in the model are:
(I_{ph}) : Light-generated current;
(I_{0}): Saturation current of the diode;
(n): Diode’s ideality factor;
(R_{s}): Equivalent series resistance;
(R_{sh}): Equivalent parallel resistance.
Characterizing the electrical parameters of PV cells is a step in optimizing their performance and reliability. This process often involves a multifaceted approach, blending theoretical analysis, numerical optimization techniques, and empirical curve fitting methods. Accurately determining the parameter values is essential for making precise predictions about a module’s power generation capabilities and identifying opportunities to enhance its efficiency through design refinements. One powerful tool employed in this parameter extraction process is optimization algorithms. These computational methods systematically adjust the model parameters until the simulated data aligns most closely with experimental measurements taken from the PV module. This iterative approach aims to converge on the optimal parameter set that minimizes the discrepancy between the modeled and observed behavior. Various optimization algorithms have been leveraged for this purpose, each with its own strengths and capabilities. Figure 3 illustrates the parameter extraction procedure used in this paper to identify the ODM parameters of the photovoltaic module.
Parameter extraction procedure.
The PV cell parameter identification and extraction scheme presents a comprehensive methodology for characterizing photovoltaic modules using a combination of measurement hardware and mathematical optimization. At the core of the setup is an Isofoten 106W-12V PV module connected to a PVPM40 IV curve tracer, which serves as the primary measurement instrument for capturing essential electrical characteristics. The measurement process involves three key parameters: voltage (({V}_{meas})), current (({I}_{meas})), and temperature (({T}_{mod})) of the PV module. These measurements are collected in real-time by the PVPM40 IV curve tracer and serve as inputs for the parameter extraction process. The system employs a one diode model as the mathematical representation of the PV cell, which includes fundamental components such as photocurrent (({I}_{ph})), a diode, series resistance (({R}_{s})), and shunt resistance (({R}_{sh})). This model effectively captures the essential electrical behavior of the PV module under various operating conditions.
where:
The parameter extraction process utilizes a sophisticated optimization approach centered around the Optimization Algorithm. This algorithm works in conjunction with a cost criteria computation system, which evaluates the difference between measured and estimated values using a specific equation (Eq. 2). The optimization process iteratively adjusts the model parameters to minimize this difference, effectively finding the best fit between the theoretical model and actual measurements. The flow of information in the system follows a logical sequence: the measured data from the curve tracer feeds into the SDM, which generates estimated outputs. These estimates are then compared with actual measurements through the cost criteria computation. The Optimization Algorithm uses this information to update the model parameters, and the process continues until optimal values are achieved. This creates a closed-loop system where the parameters are continuously refined until they accurately represent the physical characteristics of the PV module.
To validate the accuracy of our parameter extraction method, we developed an integrated simulation environment combining MATLAB™ and PSIM™ (Figs. 4 and 5). This setup allowed us to test the ODM under various environmental conditions, specifically different temperatures and irradiance levels. The model utilized the optimized parameters obtained through our optimization algorithm. The results demonstrated strong correlation between the experimental and simulated I-V curves. The optimization algorithm proved highly effective, achieving a remarkably low Root Mean Square Error (RMSE) of 0.011 amperes. This minimal error, along with the excellent parameter fit, validates both the accuracy of our optimization approach and its efficient convergence characteristics¹⁹.
The static validation model.
The dynamic validation model.
Solar power systems, while critical for the production of clean energy, deal with a number of problems that compromise their performance and sustainability. These faults can be divided into three categories: internal defects under the protective glass, exterior issues on the panel surface, and electrical failures that impact power settings. The degradation of these systems is primarily caused by four environmental factors: solar radiation exposure, which causes discoloration; temperature fluctuations, which trigger structural stress; humidity, which leads to moisture-related damage such as snail trails and connection issues; and mechanical loads from environmental forces such as wind and snow, which can cause physical damage. Understanding both fault types and their environmental sources is critical for achieving peak system performance and establishing efficient maintenance plans to extend the life of solar installations. In this context, this paper deals with short-circuit faults and partial shading detection. The experimental campaign considers 17 operating scenarios: one healthy condition (C0) and sixteen fault conditions (F1–F16), as summarized in Fig. 6. C0 corresponds to normal operation without anomalies. F1–F4 are progressive short‑circuit faults in the first string, with 2, 4, 8, and 12 series modules shorted, F5–F8 represent partial‑shading faults in the first string only, affecting 5, 9, 12, and 15 series modules. F9–F11 are combined short‑circuit faults distributed across both strings, including symmetric and asymmetric patterns. Finally, F12–F16 correspond to line‑to‑line and mixed partial‑shading faults between the two strings, such as line‑to‑line short‑circuit between specific modules and various 50–70% partial‑shading combinations across 4–8 modules in the first and second strings.
Illustration of scenarios faults.
The Wigner-Ville Distribution (WVD) is a quadratic time–frequency representation that provides optimal joint localization of signal energy^17,18. For a real-valued continuous signal (x(t) in {mathcal{L}}^{2} ({mathbb{R}})), the WVD is defined as:
where:
(t in {mathbb{R}}): Time localization parameter
(f in {mathbb{R}}): Frequency (Hz)
(tau in {mathbb{R}}): Lag variable for instantaneous autocorrelation
(xleft( t right)): Real-valued PV system measurement
The WVD belongs to Cohen’s class of time–frequency distributions and satisfies several fundamental properties that ensure its mathematical rigor and physical interpretability. First, the WVD is real-valued, i.e., (W_{x} (t,f) in {mathbb{R}}) for all (left( {t, , f} right)), which distinguishes it from complex-valued spectrograms and enables direct physical interpretation as energy density. Second, the WVD satisfies the time marginal property (energy conservation), which guarantees that integrating the time–frequency representation over all frequencies recovers the instantaneous signal energy:
Third, the frequency marginal property ensures that integrating over time yields the power spectral density, establishing consistency with classical Fourier analysis:
where (X(f) = {mathcal{F}}{ x(t)}) is the Fourier transform of the signal. Finally, the total energy preservation property confirms that the WVD conserves the total signal energy when integrated over both time and frequency domains:
These properties collectively establish the WVD as a mathematically sound time–frequency representation suitable for non-stationary PV system signal analysis.
For PV system data sampled at discrete intervals (Delta t = 1) minute with (N , = , 755) samples per condition, the discrete WVD becomes:
where:
(n in { 0,1, ldots ,N – 1}): Discrete time index
(k in { 0,1, ldots ,N – 1}): Discrete frequency bin index
(M(n) = min { n,N – 1 – n}): Maximum lag ensuring valid array indexing
The lag parameter (m) must satisfy two simultaneous constraints to ensure valid array indexing:
Taking the intersection of these constraints yields:(M(n) = min { n,N – 1 – n}).
For practical implication at signal boundaries (n = 0) or (n = N – 1), (Mleft( n right) = 0), using only the zero-lag term. At the signal center (n approx N/2), (M(n) approx N/2), utilizing maximum temporal context.
To enable uniform lag computation (M , = , N – 1) for all time indices (avoiding variable-lag computations), we extend the signal via zero-padding:
yielding (x_{{{text{ext}}}} in {mathbb{R}}^{3N – 2}).
The choice of zero-padding over alternative boundary extension methods is justified by physical considerations specific to PV system monitoring. Zero-padding represents “no signal’’ before and after the measurement window, which is physically correct for PV systems where pre-dawn and post-sunset periods have no solar input. In contrast, periodic extension would introduce artificial discontinuities by connecting sunset conditions to sunrise conditions, thereby corrupting low-frequency components associated with the diurnal cycle. Similarly, symmetric extension would create non-physical mirroring of transient events, such as cloud passages appearing to “reverse’’ in time, which contradicts the unidirectional nature of atmospheric phenomena. Therefore, zero-padding is the only boundary treatment that preserves the physical causality and finite-duration nature of solar energy conversion processes.
Given the computational complexity of full WVD ({mathcal{O}}(N^{2} )), we implement a pseudo-WVD using time-windowed FFT analysis. The signal is partitioned into (N_{{{text{time}}}} = 12) non-overlapping windows. For window (i in { 1,2, ldots ,12}), the boundaries are:
For our data (N , = , 755) and (N_{{{text{time}}}} = 12):
Window length: (L_{i} = n_{{{text{end}}}}^{(i)} – n_{{{text{start}}}}^{(i)} + 1 approx 63) samples
Temporal duration: (63 , min approx 1.05) hours per window
For each time window (i), we extract (N_{{{text{freq}}}} = 6) frequency bins using the Discrete Fourier Transform (DFT):
where (x_{i} [n] = x[n_{{{text{start}}}}^{(i)} + n]) is the signal segment in window (i).
If (L_{i} < N_{{{text{freq}}}}) (rare edge case), we zero-pad: (x_{i}^{{{text{ext}}}} [n] = left{ {begin{array}{*{20}l} {x_{i} [n],} hfill & {0 le n < L_{i} } hfill \ {0,} hfill & {L_{i} le n < N_{{{text{freq}}}} } hfill \ end{array} } right.)
The magnitude spectrum yields the pseudo-WVD: (tilde{W}_{x} [i,k] = |X_{i} [k]|).  
The computational complexity of Algorithm is significantly lower than the full WVD implementation as shown in Fig. 7. For each time window, the FFT computation requires ({mathcal{O}}(L_{i} log N_{{{text{freq}}}} ) approx {mathcal{O}}(63log 6)) operations. Aggregating over all (N_{{{text{time}}}} = 12) windows yields a total complexity of ({mathcal{O}}(N_{{{text{time}}}} cdot L_{i} log N_{{{text{freq}}}} ) = {mathcal{O}}(Nlog N_{{{text{freq}}}} )) approximate to (approx {mathcal{O}}(755log 6)). Compared to the full WVD, which requires ({mathcal{O}}(N^{2} )) operations, this represents a speedup factor of (frac{{{mathcal{O}}(N^{2} )}}{{{mathcal{O}}(Nlog N_{{{text{freq}}}} )}} approx frac{{755^{2} }}{755 times 2.6} approx 290), enabling real-time fault detection with processing times under 10 ms per sample.
Computational complexity WVD Algorithm.
Each PV sample is characterized by a 5-dimensional feature vector:
where:
(S): Solar irradiance (W/m²)
(T): Module temperature (°C)
(V_{{{text{mpp}}}}): Voltage at maximum power point (V)
(I_{{{text{mpp}}}}): Current at maximum power point (A)
(P_{{{text{mpp}}}}): Power at maximum power point (W)
Each parameter undergoes pseudo-WVD transformation independently (Algorithm WVD), producing:
In implementation, each parameter’s WVD is computed by applying time-windowed and extracting magnitude spectrum averaged to 12 time bins.
A sixth channel is created by averaging all five parameter WVDs:
This combined channel encodes holistic system behavior, capturing correlations across all parameters (e.g., simultaneous voltage drops and current spike during short-circuit faults).
Final 6 × 12 WVD matrix by stacking all channels vertically yields:
After min–max normalization (Algorithm, Step 18):
PV system monitoring has been completely transformed by the incorporation of AI into fault diagnosis and detection. AI methods offer a reliable and effective means of locating defects, guaranteeing peak performance and reducing downtime^20,21. Three important AI-based techniques for PV system problem diagnosis and detection are examined in this section.
A shallow network, commonly referred to as a classical ANN, consists of a simple architecture with a limited number of layers. It typically includes an input layer, one or two hidden layers, and an output layer. Due to their structure, shallow networks are well-suited for learning tasks involving low-dimensional data with a relatively small number of features, as they can independently extract meaningful features from the input data^21,22. Each neuron within the network processes incoming signals, performs computations, and transmits the resulting output to subsequent layers. Figure 8(a) shows a typical Classical ANN structure.
(a) A typical classical ANN design, (b) Neuron output calculation.
Each connection between neurons is associated with a numerical weight. These weights determine the strength of the connection and influence the signal transmitted between neurons. The output y of a neuron in a layer is calculated as shown in Eq. 17 and is illustrated in Fig. 8(b).
DL is regarded as the next generation of machine learning methodologies, garnering significant attention due to its advantages in pattern recognition, data mining, and knowledge discovery. Its exceptional ability to learn high-level abstract features from large datasets makes it particularly powerful. DL has been established as a viable solution for automatic database pattern recognition through the use of DNNs, which facilitate the extraction of reliable signal features for classification tasks. DNN models process multi-layer input data, training the system to achieve desired target outputs. Additionally, DNNs progressively capture increasingly abstract representations at higher layers, aligning with the hierarchical nature of these systems^23,24,25,26. The multi-layer architecture of DNNs also allows for innovative designs, such as combining supervised and unsupervised learning within a single network (see Fig. 9(a)). In supervised learning, the model leverages labeled data to establish relationships between inputs and outputs, while unsupervised learning identifies hidden patterns and structures in unlabeled data. Integrating these two approaches into a unified architecture enhances the network’s ability to handle diverse and complex challenges effectively.
(a) Typical DNN architecture; (b) sparse autoencoder structure.
A Sparse Autoencoder (SAE) is an unsupervised feature learning algorithm designed to reconstruct its inputs at the outputs while focusing on learning a compact and meaningful representation of the data. Typically, an SAE consists of three main components: an input layer, an encoder, and a decoder. The encoder transforms the input data x into a hidden representation ℎ, which contains the extracted features. The decoder reconstructs an approximation (hat{x}) of the original input based on ℎ. When the SAE is integrated into other neural networks, the decoder is often detached, and only the features extracted by the encoder are used. Figure 9(b) illustrates the architecture of a typical Sparse Autoencoder, showing the flow of data from the input layer through the hidden layer to the output layer¹⁶.
The training of an SAE involves optimizing a cost function to minimize reconstruction error and enforce sparsity. The cost function is defined as:
Here, the cost function has three main components:
Mean squared error (MSE)
Measures the reconstruction error between the input x and its reconstructed output (hat{x}).
L₂ Regularization term ((Omega_{weights})​)
Encourages smaller weights to prevent overfitting, defined as:
Sparsity regularization term ((Omega_{{{text{sparsity}}}})​)
Ensures that only a few neurons in the hidden layer are active, defined as:
In these equations:
N and K are the number of samples and inputs, respectively.
(lambda) and (beta) are coefficients for the L₂​ regularization and sparsity terms.
L represents the number of hidden layers, while (n_{l})​ and (k_{l})​ denote the output and input sizes of layer l, respectively.
(omega_{ji}^{{}})​ is the weight of the connection between neuron j and neuron i.
(KL(rho_{i} left| {hat{rho }} right._{i} )) represents the Kullback–Leibler divergence between the desired sparsity (rho_{i}) and the average activation value (hat{rho }_{i}) of neuron i.
CNNs share similarities with DNNs in their use of multiple hidden layers; however, their mathematical properties and functional behaviors differ significantly^24,27. CNNs are inspired by the way the visual cortex in the human brain processes and identifies images. These networks are primarily designed for image classification tasks. The CNN architecture consists of two main components: the feature extractor and the classifier. The feature extractor automatically learns features from the input data, while the classifier, typically implemented as a fully connected layer, performs classification based on the extracted features²⁸. A standard CNN structure comprises a sequence of alternating convolutional and pooling layers, followed by fully connected layers and a final softmax layer for output classification (see Fig. 10).
A typical CNN structure.
The convolution layer operates distinctly from other layers in a neural network. Unlike traditional layers that rely on connection weights, the convolution layer utilizes multiple kernels (or filters). Each kernel is a small matrix that slides over the input data, performing a dot product operation with sub-regions of the input. This process generates output images, referred to as feature maps.
Figure 11(a) illustrates the operation of a convolution layer. The input image is represented as (R^{3 times 4}) and the kernel is defined in (R^{2 times 2}). As the kernel moves across the input with a stride of one, it produces an output feature map of size (R^{2 times 3}).
(a) A convolution layer operation, A visual representation of the (b) mean pooling operation and (c) max pooling operation.
Pooling layers reduce the size of the image and remove redundant features. The feature map is divided into a set of subareas where each subarea is then converted to a more concise representation through a down sampling operation which can be an average pooling or a max pooling. Figures 11(b) and (c) shows a visual representation of the mean and the max pooling operations respectively.
Fully connected layers consist of interconnected neurons where the inputs from one layer are connected to every activation unit of the next layer. Fully connected layers are similar to regular neuron networks.
A Rectified Linear Unit (ReLU) layer applies a thresholding operation to modify or suppress the output values generated in the network. ReLU is one of the most widely used activation functions in image processing tasks due to its simplicity and effectiveness. It introduces non-linearity into the model, which is essential for learning complex patterns. The ReLU function is mathematically defined as:
where represents the input value. If is positive, the function returns; otherwise, it returns zero. This operation not only accelerates the convergence of deep neural networks during training but also mitigates the problem of vanishing gradients, which often occurs with other activation functions, such as the sigmoid or tanh.
The softmax layer plays a crucial role in neural networks designed for classification tasks. This layer applies the softmax function, which normalizes the exponential values of the outputs. The softmax function is differentiable and assigns a probability to each output class. It is defined mathematically as:
Here, (sigma_{i}) represents the output probability for the (i – th) class, ​ (z_{i}) is the pre-softmax output, and K denotes the total number of classes.
The modeling and characterization of PV systems are crucial for any PV system analysis. An accurate model is essential for system evaluation and simulation. Various techniques and models are available to represent a PV cell, including the single diode model, double diode model, and triple diode model. Among these, the single diode model is widely used due to its balance of simplicity and accuracy²⁹. In this study, the single diode model has been employed to replicate the characteristics of the ISOFOTON 106/12 PV module. Table 1 presents the electrical parameters of the used PV module.
The static validation of the PV module was conducted using the circuit shown in Fig. 4, which was developed and simulated using MATLAB-PSIM software. The model incorporates key environmental inputs, including solar irradiance (S) and temperature (T), as well as a variable voltage sweep ((V_{ch})​) ranging from 0 to the open-circuit voltage ((V_{oc} = 21.6{text{V}}), see Table 1). The PV module block receives these inputs to compute the voltage ((V_{pv})), current ((I_{pv})), and power ((P_{pv})​) outputs. This setup allows for the accurate generation of the I-V and P–V characteristics needed for comparison with experimental results.
The obtained results demonstrate the close alignment between experimental and simulated data for the photovoltaic (PV) module under test. Fig.12(a) illustrates the current–voltage (I-V) characteristics, showing that the simulation model accurately replicates the experimental behavior.
Obtained I-V (a) and P–V (b) characteristics under T = 27.1°C and S = 809W/m².
The short-circuit current ((I_{sc})​) is approximately 5.5 A, while the open-circuit voltage ((V_{oc})​) reaches around 21 V, with both metrics consistent between the experimental and simulated data.
Similarly, Fig. 12(b) compares the power-voltage (P–V) characteristics obtained experimentally and through simulation. The power output increases with voltage, reaching a MPP before suddenly declining as voltage approaches (V_{oc})​. Both experimental and simulated curves identify the MPP at approximately 80 W, occurring at a voltage near 18 V. This agreement highlights the model’s ability to predict the module’s maximum power output with precision. These results confirm the reliability of the simulation model in representing the non-linear characteristics of the PV module under temperature of T = 27.1°C and solar irradiance of S = 809W/m². The strong correlation between experimental and simulated data provides confidence in the model’s use for performance evaluation and optimization.
The dynamic validation of the PV system was conducted using the model illustrated in Fig. 5, implemented in the MATLAB-PSIM simulation environment. This dynamic model represents a PV array under variable environmental conditions, simulating real-world fluctuations over time. The voltage (V_{ch}) was swept dynamically from 0 to (15 times V_{oc}) to analyze the MPP performance.
Figure 13 presents the experimental and simulated results of the PV system under a clear sky day profile. Sub-Fig. 13(a) shows the current at the MPP ((I_{mpp})), Fig. 13(b) displays the voltage at the MPP ((V_{mpp})​), and Fig. 13(c) represents the power at the MPP ((P_{mpp})​). The results indicate a strong agreement between the experimental and simulated data, validating the model’s ability to dynamically capture the system’s behavior.
MPP of PV current (a), voltage (b) and power (c) under clear sky day profile experimental vs simulation.
Under the cloudy sky day profile, the dynamic performance of the PV system was evaluated by comparing experimental data with simulation results for current, voltage, and power at the MPP. Figure 14 illustrates the MPP parameters under these conditions. As shown in Fig. 14(a), (b) and (c) the measurements exhibit significant fluctuations due to the varying irradiance caused by cloud movements. Despite this, the simulated (I_{mpp}), (V_{mpp}) and (P_{mpp}) closely follows the experimental trend, validating the model’s ability to capture dynamic current variations under rapidly changing conditions.
MPP of PV current (a), voltage (b) and power (c) under cloudy sky day profile experimental vs simulation.
The results presented in this study were obtained through a detailed co-simulation approach that integrates MATLAB and PSIM environments to simulate the behavior of a PV system under both normal and faulty conditions. The simulation model incorporates dynamic environmental profiles, including temperature and solar irradiance variations for clear and cloudy days, providing realistic scenarios for evaluating system performance.
The Fig. 15(a) and (b) illustrate the impact of various fault conditions on the I-V and P–V characteristics of a PV system. Each curve represents a specific fault scenario, with the corresponding changes in (I_{pv}), (V_{pv}), and (P_{pv}) compared to the healthy PV system.
Impact of fault conditions on PV system I-V (a) and P–V (b) Characteristics.
The faults demonstrate varying degrees of impact on the PV system’s performance, with partial shading and line-to-line short circuits showing the most severe effects. Moreover, the results emphasize the importance of fault detection and mitigation strategies to maintain optimal system performance. Furthermore, the P–V curves highlight the reduction in (P_{mpp})​ under each fault scenario, demonstrating the extent to which the faults reduce the system’s power output. Finally, these findings are critical for the design of reliable monitoring systems to detect and address faults promptly.
The environmental profiles used in the simulation represent typical operating conditions. Clear day profiles exhibit smooth, continuous variations in solar irradiance (see Fig. 16), while cloudy day profiles simulate intermittent drops in irradiance caused by shading or cloud cover (see Fig. 17). Fault scenarios, such as short circuits in individual or multiple PV modules, line-to-line faults, and partial shading conditions, were integrated into the simulation model to study their impact on the system’s electrical characteristics.
Real data of solar irradiance (a) and temperature (b) under clear day.
Current (a), Voltage (b) and Power (c) at MPP under clear sky day within heathy and faulty PV system.
The results are derived using an MPP tracking algorithm, which continuously adjusts the PV system’s operating point to calculate the maximum power point parameters under varying conditions. The obtained data include time-series plots of (I_{mpp}), (V_{mpp}) and (P_{mpp})​ for healthy and faulty conditions. These results highlight the system’s performance degradation under specific faults and provide insights into the influence of environmental variations on fault behavior.
The temperature (see Fig. 16(a)) increases during the day, peaks around midday, and decreases towards the evening. It exhibits fluctuations that might be attributed to environmental changes. On the other hand, solar irradiance (see Fig. 16(b)) follows a smooth curve, reaching its peak at midday when the sun is at its highest point in the sky. The shape of the curve aligns with the expected solar path under clear sky conditions.
The PV system under normal operating conditions exhibits optimal performance. The current, voltage, and power at the MPP follow smooth curves, with the power peaking around midday when solar irradiance is highest. These results serve as the baseline for comparison in Fig. 17. While, the short circuit of a single module has a minor impact on the overall system performance. The current and voltage at MPP experience slight reductions, leading to a small decrease in power output. This indicates that the system is relatively resilient to isolated faults, as other modules compensate for the loss. Moreover, a short circuit affecting ten modules causes a significant drop in system performance. The current at MPP decreases notably, as the faulty modules no longer contribute to the total current. The voltage at MPP is also significantly reduced due to the loss of the series-connected modules. Consequently, the power output at MPP is substantially lower than the healthy condition, demonstrating the severity of this fault. However, 5 × 10 modules Line-to-Line short-circuited is the most severe fault scenario analyzed. The line-to-line short circuit involving five sets of ten modules results in an unstable and drastically reduced performance. The current output is the lowest among all scenarios, as large portions of the array are bypassed. The voltage curve becomes unstable and highly reduced, further exacerbating the loss of power output. This scenario highlights the critical impact of widespread faults on the PV system’s efficiency and stability. Finally, partial shading of five modules at 50% leads to a moderate decrease in system performance. The current at MPP is reduced due to the lower current contribution from the shaded modules, while the voltage at MPP experiences a slight decrease. The overall power output is affected but remains considerably higher than in scenarios involving severe faults. This result underscores the mismatch losses caused by partial shading and the importance of shading mitigation strategies, such as bypass diodes or reconfiguration techniques.
Each fault has a unique impact on the PV system’s performance, with the severity of the impact increasing with the number of affected modules and the type of fault. While the system shows resilience to isolated faults, extensive short circuits and line-to-line faults severely degrade its efficiency and stability. Partial shading introduces mismatch losses, highlighting the importance of implementing fault-tolerant designs and shading mitigation strategies to enhance the reliability and performance of PV systems.
This work uses both clear and cloud days for faults detection and diagnosis the cloudy day is presented in Fig. 18. The analysis of solar irradiance and temperature measurements during a cloudy day reveals significant temporal variations and correlations between these two parameters. The data exhibits notable characteristics of cloudy conditions, manifested through irregular, jagged patterns in both measurements, indicating the influence of intermittent cloud cover.
Real data of solar irradiance (a) and temperature (b) under cloudy day.
A healthy PV system is the standard, demonstrating consistent and optimal performance across all parameters. Short circuits, especially those involving multiple modules or line-to-line faults, lead to substantial current increases, abrupt voltage decreases, and considerable power loss as presented in Fig. 19. These faults signify significant system disruptions and an elevated risk of failure. Partial shading consistently reduces current, voltage, and power, indicating efficiency losses while avoiding total system failure.
Current (a), Voltage (b) and Power (c) at MPP under cloudy sky day within heathy and faulty PV system.
In this study, a dedicated experiment employs the WVD to convert numerical PV system data into image-like representations tailored for CNN-based fault detection and diagnosis, as illustrated in Fig. 20. The input dataset comprises five critical parameters–-solar irradiance, temperature, voltage at MPP, current at MPP, and power at MPP, which are initially arranged into a feature matrix of size (12835 times 5)((samples , times features)), obtained by concatenating 755 samples for each of the 17 operating conditions (one healthy state and sixteen fault states) under clear- and cloudy-sky profiles.
Transformation of multi-parameter PV signals to time–frequency images via WVD.
Before data-quality filtering, an index split is applied, yielding 6418 training samples and 6417 testing samples, each represented by a 5-dimensional feature vector. An irradiance threshold of ((S ge 100 {text{W/m}}^{2} )) is then imposed to remove low-irradiance points; this step discards 1947 training and 1946 testing samples, retaining 4471 samples in each subset and resulting in a final post-filtering dataset of 8942 samples 4471 training, 4471 testing, stored as (X_{{{text{train}}}} in {mathbb{R}}^{4471 times 5}) and (X_{{{text{test}}}} in {mathbb{R}}^{4471 times 5}). For the WVD-based CNN experiment, each one-dimensional signal in these feature matrices is transformed into a fixed-size (6 times 12) time–frequency image, and the resulting images are stacked into a four-dimensional array of size (6 times 12 times 1 times N), where the first two dimensions encode the time–frequency map, the third dimension denotes a single channel, and the fourth dimension indexes the (N) retained samples used for CNN training and evaluation.
The WVD transformation reveals distinctive patterns across all PV operating conditions (see Fig. 21), producing a unique time–frequency ‘fingerprint’ for each of the sixteen fault classes F1–F16 compared with the healthy baseline C0. In the healthy state, the WVD maps exhibit a uniform energy distribution over time for all parameters ((I_{mpp}), (V_{mpp}) and (P_{mpp})), with smooth, stationary behaviour in the combined row, establishing a reference signature for normal operation. For progressive short‑circuit faults in the first string (F1–F4: SC of 2, 4, 8, and 12 modules), the WVD images show increasingly pronounced disturbances in the current and power rows, accompanied by localized changes in the voltage pattern, while the irradiance and temperature rows remain close to the healthy profile; the severity of the fault is reflected in the growth and spatial extension of high‑energy regions. Pure partial‑shading faults in the first string (F5–F8: shading of 5, 9, 12, and 15 modules) introduce characteristic non‑uniform energy bands in the irradiance row and correlated distortions in Pmpp , clearly differentiating them from short‑circuit faults despite comparable reductions in output power. The combined faults affecting both strings (F9–F16) generate more complex WVD signatures: mixed short‑circuit patterns (F9–F11) produce asymmetric energy concentrations between current and voltage rows of the two strings, whereas line‑to‑line and mixed partial‑shading cases (F12–F16) lead to simultaneous distortions in the irradiance, temperature, and power rows, reflecting strong coupling between electrical and environmental effects. Overall, these results confirm that the WVD effectively converts multi‑parameter PV measurements into discriminative 6 × 12 time–frequency images, where each operating class (C0–F16) is associated with a distinctive energy distribution, providing a rich feature space for CNN‑based classification of complex PV system behaviours.
Features extraction from WVD transformation for PV system under faulty conditions.
This transformation enables the CNN to effectively process and analyze the data by learning spatial features from the time–frequency representations of the PV parameters. By converting abstract numerical signals into visually interpretable image-like formats, the WVD enhances the CNN’s ability to identify intricate patterns and relationships within the PV system’s operational characteristics. The result is a framework capable of extracting meaningful features, supporting advanced analysis and predictive modeling for PV systems. This approach bridges the gap between time–frequency analysis and deep learning, providing a robust methodology for handling complex PV data in an interpretable and computationally efficient manner.
In order to verify the accuracy of the AI models, simulations were performed using real data. These simulations were carried out in two different settings: one with no background noise and another with added noise. In the noiseless environment, measurements are inherently free of noise. Each measurement device had its output supplemented with random numbers from a specified range to simulate a noisy environment. In this work an exaggerated noisy has been applied to test the accuracy of faults detection and diagnosis the amplitude of noisy attributes has chosen as follows:
In this work, the network consists of an input layer, an output layer, and 2 hidden layers. The input layer has 5 neurons (left[ {S,T,V_{mp} ,I_{mp} ,P_{mp} } right]). The output layer is a softmax layer that has 5 neurons each representing the probability of each class (from 1 to 5) respectively. The two hidden layers consists of 20 neurons each where their transfer functions are hyperbolic tangent sigmoid ((tansig)) and linear ((purelin)) respectively as illustrated in Fig. 22. The network has been trained using the scaled conjugate gradient (SCG) algorithm and its performance has been measured by computing the cross-entropy loss between predictions and targets as given in Eq. 28.
where (y_{ni}) is the desired neural network output (target which is the class from 1 to 5), (hat{y}_{ni}) is the neural network output, (N) and (K) being the numbers of observations, and classes, respectively.
Classical ANN structure for fault detection.
The proposed system combines unsupervised and supervised deep learning techniques to enhance fault detection in photovoltaic systems. The architecture consists of two distinct stages: feature extraction and fault classification. In the feature extraction phase, a stacked sparse autoencoder (SAE) is employed to autonomously learn and improve significant representations from the input data (see Fig. 23(a)). An initial autoencoder condenses five essential photovoltaic parameters into a representation comprising 12 features. The encoding process employs an encoder layer incorporating a saturating linear transfer function, whereas the decoding phase reconstructs the original inputs through a purely linear function. The second autoencoder processes the 12 extracted features, encoding them into a representation of 36 refined features via a similar encoding and decoding process (see Fig. 23(b)). The 36 refined features are input into a classical supervised ANN for fault classification as proposed in Fig. 23(c). Figure 23(d) illustrates the complete end-to-end architecture of the proposed fault detection system for photovoltaic (PV) systems, integrating both feature extraction and fault classification stages into a unified framework. The input layer consists of key PV parameters. These inputs are processed through a series of layers for feature extraction and classification.
(a) First autoencoder architecture for feature extraction, (b) Second autoencoder refining features, (c) Supervised ANN for fault classification and (d) End-to-end fault detection neural network architecture.
The ANN associates the extracted features with predefined categories via a multi-class output layer, including healthy, shaded, or faulty PV conditions. This hybrid architecture facilitates effective dimensionality reduction while preserving essential information, thus allowing for precise and dependable fault diagnosis in complex or noisy settings. Integrating unsupervised autoencoders for feature extraction with a supervised artificial neural network for classification improves the adaptability and robustness of the system, presenting a viable solution for photovoltaic system monitoring and diagnostics.
The proposed CNN architecture, detailed in Table 2, is specifically designed to extract discriminative features from (6 times 12 times 1) WVD time–frequency representations and accurately classify 17 distinct PV system fault conditions. The network comprises 18 layers organized into a hierarchical structure: input normalization, two convolutional feature extraction blocks, two fully connected classification blocks, and a softmax output layer for multi-class probability estimation. The architecture begins with an input layer (L0) that processes (6 times 12 times 1) WVD images, where spatial dimensions encode temporal evolution (12 time bins) and frequency content (6 frequency bins) of electrical signals from PV arrays. Input data undergoes z-score normalization to ensure zero mean and unit variance, accelerating convergence and stabilizing gradient flow during training.
Convolutional Block 1 (Layers L1-L4) extracts low-level spatial features through a (3 times 3) convolutional layer with 32 filters (L1), producing (6 times 12 times 32) feature maps. Same padding preserves spatial dimensions, preventing information loss at boundaries. Batch normalization (L2) standardizes activations across mini-batches with epsilon (10^{ – 5}) and momentum 0.9, reducing internal covariate shift. ReLU activation (L3) introduces non-linearity via (f(x) = max (0,x)), enabling the network to learn complex decision boundaries. Max pooling (L4) with a (2 times 2) kernel and stride 1 performs controlled spatial down-sampling to (5 times 11 times 32), achieving translation invariance while retaining dominant fault signatures. Convolutional Block 2 (Layers L5-L8) builds hierarchical representations by applying 64 filters with (3 times 3) kernels (L5), doubling representational capacity to capture subtle inter-class differences among the 17 fault categories. The increased filter count enables detection of complex patterns such as multi-frequency harmonics (partial shading), sustained low-frequency components (line-to-line faults), and transient spikes (open-circuit conditions). Following the same normalization-activation-pooling sequence (L6-L8), this block produces (4 times 10 times 64) high-level feature maps, compressing spatial information while amplifying semantic content essential for fault discrimination. The fully connected classification blocks transform convolutional features into class predictions. After flattening (L9) the (4 times 10 times 64) feature maps into a 2,560-dimensional vector, two successive fully connected layers with 128 neurons (L10) and 64 neurons (L13) progressively compress the representation into increasingly abstract feature spaces optimized for multi-class separation. Each FC layer is followed by ReLU activation (L11, L14) and dropout regularization (L12: 30%, L15: 20%) during training. Dropout randomly deactivates neurons to prevent co-adaptation, forcing the network to learn robust, redundant representations that generalize beyond the training set. During inference, dropout is disabled, and all neurons contribute to predictions. The output layer (L16-L18) maps the 64-dimensional feature vector to 17 class logits via a fully connected layer (L16). The softmax activation (L17) converts logits into a probability distribution (widehat{{mathbf{y}}} in {mathbb{R}}^{17}) where (sumlimits_{j = 1}^{17} {hat{y}_{j} } = 1), with each (hat{y}_{j}) representing the probability of fault class (j). The classification layer (L18) applies cross-entropy loss to measure discrepancy between predicted probabilities and true labels, guiding backpropagation-based weight optimization. The predicted fault class is determined by (hat{c} = arg max_{j} hat{y}_{j}). Figure 24 demonstrates the systematic transformation of input data through successive layers, highlighting dimensionality evolution ((6 times 12 times 1 to 6 times 12 times 32 to 5 times 11 times 64 to 4 times 10 times 64 to 2560 to 128 to 64 to 17)) and functional specialization at each stage. The multi-stage regularization strategy (batch normalization, dropout, L2 weight decay) ensures reliable generalization to unseen fault patterns, making the architecture suitable for real-time deployment in large-scale PV monitoring systems.
CNN structure for fault detection.
The evaluation of the three AI models—classical ANN, DNN, and CNN—was conducted under noiseless conditions to assess their effectiveness in fault detection and classification in the absence of external noise influencing the measurements. The classical ANN classifier exhibited perfect performance in this optimal scenario, attaining 100% accuracy as presented in Fig. 27(a). This outcome highlights the model’s capacity to learn and categorize patterns efficiently within a controlled setting characterized by clean data and predominant linear relationships. However, the performance is constrained by the simplicity of its architecture, which may not effectively generalize to more complex data scenarios. The DNN classifier (see Figs. 25(b), 26) demonstrated high performance, attaining an accuracy rate of 98.5%. The stacked sparse autoencoder architecture facilitates extracting hierarchical features from the input data, enhancing its capability to identify subtle patterns compared to the Classical ANN. The marginal accuracy difference in noiseless conditions indicates that the DNN’s capacity to capture complex relationships may offer significant advantages when applied to more intricate datasets. The CNN classifier demonstrated superior performance to the other models, achieving an accuracy score of 99.3% (Fig. 27(c)). The CNN utilized convolutional and pooling layers to extract advanced spatial features, showcasing its enhanced capability to process structured data representations. The preprocessing step converts numerical inputs into image-like representations utilizing the Wigner-Ville distribution, enhancing CNN’s performance by supplying a more comprehensive feature set for classification.
Confusion matrix of the test data under noiseless measurements.
Confusion matrix of the test data under noisy measurements.
The utilized daily profile of real measured at MPP of voltage, current and power.
This Table 3 highlights how each neural network processes data differently, reflecting the differences in their architectures and the sophistication of their feature extraction methods.
All three models exhibited strong performance in noiseless conditions, with the CNN demonstrating a slight advantage over the others. The findings demonstrate that training under noiseless conditions creates an optimal environment for artificial neural networks to enhance their classification accuracy. Classical ANN can achieve perfect accuracy under certain conditions; however, the advanced architectures of DNN and CNN provide enhanced flexibility and robustness for handling more complex data patterns, even in optimal scenarios.
The Classical ANN classifier showed significant limitations when exposed to noisy data. Its overall accuracy dropped drastically to 51.3% as shown in Figs. 25(a), 26, highlighting its inability to generalize well in the presence of noise. The absence of advanced feature extraction mechanisms in its architecture makes the Classical ANN susceptible to noise, causing inconsistent predictions. While it remains computationally efficient, its performance under noisy conditions is inadequate for practical applications. The DNN classifier demonstrated considerable resilience to noise, maintaining an overall accuracy of 85.1% is presented in Fig. 25(b), 26. The stacked sparse autoencoders within the DNN allowed the model to extract robust features from noisy data, enabling it to identify underlying patterns despite the perturbations. This hierarchical feature extraction mechanism provided the DNN with a significant advantage over the Classical ANN in handling complex, noise-affected datasets. However, the model’s computational demands were notably higher due to the additional feature extraction stages. The CNN classifier (see Fig. 25(c), 26) achieved the best performance among the three models in noisy conditions, with an overall accuracy of 86.0%. By leveraging its convolutional and pooling layers, the CNN effectively extracted spatial features from the image-like representations of the data, mitigating the impact of noise. The Wigner-Ville distribution preprocessing step further enhanced the robustness of the CNN by transforming noisy numerical inputs into structured representations, which the model could process more efficiently. Despite the additional computational overhead introduced by preprocessing, the CNN’s consistent and accurate predictions make it the most reliable choice for noisy environments.
The results under noisy measurements highlight the limitations of the Classical ANN and the superiority of DNN and CNN models in handling noise. While the DNN offers robust feature extraction and strong performance, the CNN slightly outperforms it due to its advanced spatial feature extraction and noise resilience. For real-world applications with noisy data, the CNN is the preferred choice, followed closely by the DNN, with the Classical ANN being unsuitable for such conditions.
In this subsection we present the results of classification accuracy and true positive rate under noisy attributes for all classifiers using a daily profile. The Fig. 27 illustrates a solar energy system’s maximum power output across different operational conditions and the results of classifiers in Figs. 28 and 29, emphasizing the system’s performance in both optimal and malfunctioning states. After heathy state 16 states are indicate distinct fault scenarios, each exhibiting diminished voltage, current and power output relative to the healthy state.
Faults detection under noiseless data.
Faults detection under noisy data.
Under noiseless conditions in Fig. 28, all three architectures achieve very high overall accuracy: about 99.5% for the ANN, 99.5% for the DNN, and 97.1% for the CNN. Each subplot shows a compact cluster of colored markers for every fault class along the sample’s axis, with almost no dispersion. This indicates that nearly all test samples in each of the 17 PV operating states (healthy plus 16 faults) are correctly classified, and misclassifications are extremely rare for every class. The nearly identical positions of the ANN and DNN clusters suggest that the deeper, autoencoder-based representation brings only marginal gains over a well-tuned shallow ANN when measurement noise is absent. The CNN remains slightly less accurate overall but still maintains tight, well-separated clusters for all classes, confirming that the proposed time–frequency features are highly discriminative in ideal conditions. When realistic noise is added to irradiance and temperature measurements, overall accuracy drops to roughly 82.2% for the ANN and 76.3% for the DNN, while the CNN still reaches about 90.3%. In the noisy Fig. 29, several fault classes exhibit scattered markers for ANN and especially DNN, meaning that many samples deviate from their ideal class cluster and are more prone to misclassification. This effect is most visible for intermediate fault states, where overlap between neighboring classes becomes pronounced. In contrast, the CNN markers remain comparatively compact and well aligned across most classes, indicating that the convolutional architecture, operating on structured time–frequency inputs, is more robust to measurement perturbations and preserves class separability better than the fully connected models. Taken together, the noiseless and noisy results show that all three models are essentially perfect fault detectors in clean conditions, but their robustness diverges once environmental volatility is introduced. The ANN and DNN provide excellent accuracy when noise is negligible, whereas the CNN offers a more favorable accuracy–robustness trade-off, retaining the highest overall accuracy and the most stable per-class behavior under noisy conditions. For practical PV monitoring, where sensor noise and irradiance/temperature fluctuations are unavoidable, these figures support choosing the CNN-based implementation as the preferred model, while recognizing that ANN/DNN remain competitive baselines in controlled environments.
The bar chart in Fig. 30 summarizes how seven classification models perform under noiseless and noisy conditions in terms of overall accuracy for PV fault detection. Under noiseless conditions, the three deep learning models achieve the best results: the classical ANN and the DNN both reach accuracies around 99.5–99.6%, while the CNN attains about 97.1% accuracy. Among the conventional machine learning methods, Random Forest follows with approximately 93.6% accuracy, then SVM at about 88.6%, GBM at roughly 84%, and kNN at around 75.7%. These values show that when measurements are clean, all neural approaches are highly reliable, clearly outperforming the tree-based, margin-based, and distance-based baselines.​ When noise is introduced into the test data, the accuracy of every model decreases, but the magnitude of this degradation differs substantially between methods. The CNN becomes the most robust architecture, maintaining an accuracy of about 90.3%, which represents the smallest relative drop among all models. In contrast, the ANN and DNN fall to roughly 82.2% and 76.3%, respectively, indicating that fully connected architectures are more sensitive to measurement perturbations. For the classical ML models, Random Forest and SVM remain comparatively resilient, with accuracies near 82.8% and 83.9%, whereas GBM and kNN decline more markedly to around 73.9% and 72.8%. Overall, the figure suggests that while deep models (especially ANN/DNN) are optimal in ideal, low-noise settings, CNN and, to a lesser degree, RF and SVM offer a better trade-off between accuracy and robustness for real-world PV monitoring where noisy sensor data are unavoidable.
Models comparisons under noisy and noiseless.
Under noiseless attributes (Table 4), the deep models (CNN, DNN, ANN) achieve near-perfect precision and recall across almost all classes, with overall accuracy values around 97–99%, whereas GBM, RF, SVM, and especially kNN show lower and more variable true positive rates for several faults. Under noisy attributes (Table 5), all models experience a degradation in both precision and recall, but CNN, RF and SVM retain relatively high per-class values and the best accuracies, indicating stronger robustness to measurement noise than DNN, ANN, GBM and kNN.
To validate deployment readiness, we systematically evaluated all seven classification methods under varying noise conditions representative of sensor aging and environmental interference. Table 6 presents accuracy across seven noise levels (0 times) to (3 times) baseline, while last raw quantifies degradation rates via linear regression. CNN achieves the optimal balance between high baseline accuracy (97.1%) and competitive noise robustness (8.91% unit degradation), maintaining 90.8% accuracy at (1 times) noise—the operational condition expected in properly maintained installations—and 71.5% accuracy at (3 times) noise, representing end-of-life sensor performance or extreme electromagnetic interference. This robustness stems from CNN’s architectural noise averaging: the (6 times 12) WVD matrix aggregates (sim 63) temporal samples per pixel, yielding (sqrt {63} approx 7.9 times) noise variance reduction ({text{Var}}[W[i,j]] = {raise0.5exhbox{$scriptstyle {sigma_{{{text{input}}}}^{2} }$} kern-0.1em/kern-0.15em lower0.25exhbox{$scriptstyle {L_{{{text{window}}}} }$}} approx {raise0.5exhbox{$scriptstyle {sigma_{{{text{input}}}}^{2} }$} kern-0.1em/kern-0.15em lower0.25exhbox{$scriptstyle {63}$}}), complemented by 2D convolutional spatial smoothing.
The noise stability analysis (Table 6) reveals a fundamental trade-off in fault detection systems: methods optimized for noiseless accuracy (ANN, DNN: 99.5%) exhibit catastrophic degradation under operational noise 52–53% at (3 times), while methods with lower baseline performance (kNN: 75.7%) degrade gracefully (57.5% at (3 times), degradation rate 6.20% unit). CNN uniquely achieves near-optimal performance on both axes: competitive noiseless accuracy (97.1%, within 2.4% of ANN/DNN) and superior noisy accuracy (71.5% at (3 times)), + 14% absolute advantage over nearest competitor. This validates the WVD + CNN paradigm’s suitability for real-world PV monitoring, where sensor noise is not an experimental artifact but an operational reality. For installations with 3–5 year sensor recalibration intervals, CNN maintains actionable accuracy (> 80% ) across the sensor’s operational lifetime, whereas competing methods would require more frequent maintenance or accept degraded fault detection reliability.
To provide a safety–oriented assessment beyond raw accuracy, the proposed framework is evaluated using a comprehensive suite of multi–class metrics computed from the confusion matrices of all seven models (Classical ANN, DNN with stacked autoencoders, CNN, GBM, RF, SVM, and kNN) under both noiseless and noisy operating conditions. For each fault class $k$, we derive true positives, false negatives, false positives, and true negatives from the confusion matrix and compute precision, recall (sensitivity), and specificity; these quantities are then macro–averaged over all (K = 17) classes to explicitly account for the class–imbalance between healthy operation and the different PV fault types. On top of these class–wise measures, we report the macro–averaged F1–score, Cohen’s (kappa), and balanced accuracy, together with macro sensitivity and specificity, in two new tables summarizing the performance of all seven models for noiseless and noisy conditions, respectively.
To provide a safety–oriented assessment beyond raw accuracy, the proposed framework is evaluated using a comprehensive suite of multi–class metrics computed from the confusion matrices of all seven models (Classical ANN, DNN with stacked autoencoders, CNN, GBM, Random Forest, SVM, and kNN) under both noiseless and noisy operating conditions. For each fault class (k), we derive true positives, false negatives, false positives, and true negatives from the confusion matrix and compute precision, recall (sensitivity), and specificity; these quantities are then macro–averaged over all 17 classes to explicitly account for the class–imbalance between healthy operation and the different PV fault types. On top of these class–wise measures, we report the macro–averaged F1–score, Cohen’s (kappa), and balanced accuracy, together with macro sensitivity and specificity, in two tables summarizing the performance of all seven models for noiseless and noisy conditions, respectively.
Let (C) be the confusion matrix with entries (C_{ij})(true class (i), predicted class (j)) and (K = 17) classes. For each class (k),
Per-class precision, sensitivity, and specificity are :
We use macro-averaging to reduce the influence of class imbalance:
The macro F1-score is (F1 = frac{1}{K}sumlimits_{k} {frac{{2{text{Prec}}_{k} {text{Sens}}_{k} }}{{{text{Prec}}_{k} + {text{Sens}}_{k} }}}). and balanced accuracy is ({text{Balanced Accuracy}} = tfrac{1}{2}({text{Sensitivity}} + {text{Specificity}}).) Cohen’s (kappa) is computed from the observed accuracy (p_{o} = frac{1}{N}sumlimits_{k} {C_{kk} } {text{ and }}N = sumlimits_{i,j} {C_{ij} } ,) and the expected chance agreement (p_{e} = frac{1}{{N^{2} }}sumlimits_{k} ( sumlimits_{j} {C_{kj} } )(sumlimits_{i} {C_{ik} } ),) Via (kappa = frac{{p_{o} – p_{e} }}{{1 – p_{e} }}.)
From Table 7 the deep models (ANN, DNN, CNN) achieve macro F1 above 97% and (kappa > 0.96), indicating near-perfect agreement with the ground truth and very limited class-imbalance bias, whereas the classical ML baselines (GBM, RF, SVM, kNN) show noticeably lower balanced accuracy and sensitivity. Under noisy irradiance and temperature perturbations (Table 8), all models degrade, but ANN, CNN, RF, and SVM maintain (kappa > 0.80) and specificity above 98%, demonstrating robustness to environmental volatility while still limiting false.
This study presents a comprehensive evaluation of fault detection and classification in photovoltaic (PV) systems using artificial intelligence, systematically comparing seven classifiers—classical Artificial Neural Network (ANN), Deep Neural Network (DNN), Convolutional Neural Network (CNN), Gradient Boosting Machine (GBM), Random Forest (RF), Support Vector Machine (SVM), and k-Nearest Neighbors (kNN)—all trained on structured Wigner-Ville Distribution (WVD) time–frequency representations. The experimental campaign encompasses 17 operating conditions, including one healthy state (C0) and 16 fault types (F1–F16) spanning progressive short-circuit faults, partial-shading scenarios, and combined inter-string fault patterns, providing a rigorous testbed for evaluating diagnostic performance across diverse fault signatures.​
Under noiseless conditions, the classical ANN and DNN achieved the highest accuracies of 99.51% and 99.49%, respectively, demonstrating their strong capacity to learn discriminative patterns when measurement uncertainty is minimal. The CNN attained 97.09% accuracy in ideal conditions, while tree-based and instance-based methods showed progressively lower performance: RF (93.47%), SVM (88.62%), GBM (84.01%), and kNN (75.69%). These results establish that fully-connected neural architectures excel when data quality is high and feature relationships are consistent.​
However, under realistic sensor noise simulating operational variability (± 10 W/m² irradiance, ± 2°C temperature, ± 0.5 V voltage, ± 0.2 A current, ± 5 W power at baseline), the performance hierarchy fundamentally shifts. The CNN demonstrated the highest noise robustness with 90.27% accuracy, significantly outperforming SVM (83.85%), RF (82.80%), ANN (82.20%), DNN (76.27%), GBM (73.85%), and kNN (72.80%). This superior performance under adverse conditions reflects the CNN’s ability to extract hierarchical spatial features from structured WVD time–frequency images, combined with multi-stage regularization (batch normalization, dropout, L2 weight decay) that enhances generalization. Quantitative degradation analysis reveals that the CNN degrades at only 8.91% per noise unit compared to 16.27% for ANN, 15.00% for DNN, and 11.34% for RF, confirming its optimal accuracy-robustness trade-off for long-term deployment in real-world PV installations where sensor drift and environmental perturbations are unavoidable.​
The integration of WVD transformation with CNN architecture addresses critical limitations of traditional fault detection methods. By converting multivariate electrical measurements into compact 6 × 12 time–frequency images, the framework enables automated feature extraction without manual engineering, while the pseudo-WVD implementation achieves computational complexity of O(NM log M) with processing times under 10 ms per sample, making real-time monitoring feasible. The structured representation preserves both temporal dynamics and frequency characteristics of fault signatures, providing superior discriminative power compared to raw electrical parameters or single-domain features.​
While the proposed CNN-WVD framework demonstrates strong performance under controlled experimental conditions, several important limitations must be acknowledged. First, training requires GPU acceleration for efficient WVD transformation and optimization of the 18-layer architecture, though inference can execute on standard industrial CPUs. Second, the model was trained on a specific 30-panel installation in Oran, Algeria, and site-specific fine-tuning or transfer learning may be required when deploying to installations with different climatic conditions, PV module technologies, or mounting configurations. Third, the framework currently relies exclusively on electrical measurements without integrating complementary imaging modalities such as infrared thermography or electroluminescence. Fourth, although noise robustness analysis incorporates uncertainty levels up to 3 × baseline sensor error to simulate aging effects, the framework has not undergone validation in full-scale operational environments over extended periods. These limitations define a clear roadmap for advancing PV fault diagnostics toward fully autonomous, scalable, and generalizable real-world deployment through pilot field studies, multimodal sensor fusion, incremental learning strategies, and cross-site validation protocols.​
The advantages of the CNN-WVD framework notably outweigh the drawbacks, enabling more reliable and proactive PV maintenance through high accuracy, superior noise resilience, and automation capabilities. The computational demands and training scope requirements can be managed through engineering solutions (cloud-based training infrastructure, edge computing for inference) and further research into transfer learning and few-shot adaptation. As global PV installations continue to expand, intelligent fault detection and diagnosis systems will become increasingly vital for maximizing energy yield, reducing downtime, and supporting the transition to sustainable energy systems. The comprehensive benchmarking, rigorous mathematical framework, and transparent discussion of limitations presented in this work provide a solid foundation for future advances in AI-driven PV diagnostics and their practical implementation in operational solar power plants.
The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request.
Sizkouhi, A. M., Aghaei, M. & Esmailifar, S. M. A deep convolutional encoder-decoder architecture for autonomous fault detection of PV plants using multi-copters. Sol. Energy 223, 217–228 (2021).
Article  ADS  Google Scholar 
Lu, X. et al. Efficient fault diagnosis approach for solar photovoltaic array using a convolutional neural network in combination of generative adversarial network under small dataset. Sol. Energy 253, 360–374 (2023).
Article  ADS  Google Scholar 
Latoui, A. & Daachi, M. E. H. Real-time monitoring of partial shading in large PV plants using Convolutional Neural Network. Sol. Energy 253, 428–438 (2023).
Article  ADS  Google Scholar 
Yousif, H. & Al-Milaji, Z. Fault detection from PV images using hybrid deep learning model. Sol. Energy 267, 112207 (2024).
Article  Google Scholar 
Chen, X., Gao, W., Hong, C. & Tu, Y. A novel series arc fault detection method for photovoltaic system based on multi-input neural network. Int. J. Electr. Power Energy Syst. 140, 108018 (2022).
Article  Google Scholar 
Lu, X. et al. Fault diagnosis model for photovoltaic array using a dual-channels convolutional neural network with a feature selection structure. Energy Convers. Manag. 248, 114777 (2021).
Article  Google Scholar 
Ramírez, I. S., Márquez, F. P. G. & Chaparro, J. P. Convolutional neural networks and Internet of Things for fault detection by aerial monitoring of photovoltaic solar plants. Measurement https://doi.org/10.1016/J.MEASUREMENT.2024.114861 (2024).
Article  Google Scholar 
Qu, J., Sun, Q., Qian, Z., Wei, L. & Zareipour, H. Fault diagnosis for PV arrays considering dust impact based on transformed graphical features of characteristic curves and convolutional neural network with CBAM modules. Appl. Energy 355, 122252 (2024).
Article  Google Scholar 
Hong, Y. Y. & Pula, R. A. Detection and classification of faults in photovoltaic arrays using a 3D convolutional neural network. Energy 246, 123391 (2022).
Article  Google Scholar 
Kellil, N., Aissat, A. & Mellit, A. Fault diagnosis of photovoltaic modules using deep neural networks and infrared images under Algerian climatic conditions. Energy 263, 125902 (2023).
Article  Google Scholar 
Hou, Z., Liu, J., Shao, Z., Ma, Q. & Liu, W. Machine learning innovations in renewable energy systems with integrated NRBO-TXAD for enhanced wind speed forecasting accuracy. Electronics 14, 2329 (2025).
Article  Google Scholar 
Hu, Q., Fan, C. & Guo, D. Forecasting future power via machine learning models for predicting energy consumption. Sādhanā 50(3), 225 (2025).
Article  Google Scholar 
Liu, J., Hou, Z. & Yin, T. Short-term power load forecast using OOA optimized bidirectional long short-term memory network with spectral attention for the frequency domain. Energy Rep. 12, 4891–4908 (2024).
Article  Google Scholar 
Hou, Z., Wang, B., Liu, J., He, Y. & Yao, Y. Physics-inspired time-frequency feature extraction and lightweight neural network for power quality disturbance classification. Front. Phys. 13, 1616367 (2025).
Article  Google Scholar 
Hou, Z., Liu, J. & Yu, S. Enhanced analog circuit fault diagnosis via continuous wavelet transform and dual-stream convolutional fusion. Sci. Rep. 15(1), 19828 (2025).
Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 
Seghiour, A., Abbas, H. A., Chouder, A. & Rabhi, A. Deep learning method based on autoencoder neural network applied to faults detection and diagnosis of photovoltaic system. Simul. Model Pract. Theory 123, 102704 (2023).
Article  Google Scholar 
Garoudja, E., Chouder, A., Kara, K. & Silvestre, S. An enhanced machine learning based approach for failures detection and diagnosis of PV systems. Energy Convers. Manag. 151, 496–513 (2017).
Article  ADS  Google Scholar 
Chouder, A., Silvestre, S., Taghezouit, B. & Karatepe, E. Monitoring, modelling and simulation of PV systems using LabVIEW. Sol. Energy https://doi.org/10.1016/j.solener.2012.09.016 (2013).
Article  Google Scholar 
Elmamoune, H., Abdellatif, S., Lakhdar, M., Aissa, C. & Said, B. An Improved Grey Wolf Optimizer for Parameter Extraction of Photovoltaic Cells. https://www.icaens.com/.
Malik, A., Haque, A., Kurukuru, V. S. B., Khan, M. A. & Blaabjerg, F. Overview of fault detection approaches for grid connected photovoltaic inverters. e-Prime – Adv. Electr. Eng., Electron. Energy 2, 100035 (2022).
Article  Google Scholar 
Saifi, I. A., Amir, M., Haque, A. & Iqbal, A. Investigation of condition monitoring system for grid connected photovoltaic (GCPV) system with power electronics converters using machine learning techniques. e-Prime – Adv. Electr. Eng., Electron. Energy 9, 100722 (2024).
Article  Google Scholar 
Seghiour, A., Seghier, T., Zegnini, B. & Georgoulas, G. Multi-Class Classification Approach for the Diagnosis of Broken Rotor Bars Based on Air-Gap Magnetic Flux Density Spectrum. Electrotehnica, Electronica. Automatica (EEA) 65, 31–39 (2017).
Google Scholar 
El Fadili, Y. & Boumhidi, I. New design of an intelligent electromagnetic torque controller based on neural network and fractional calculus: Variable-speed wind energy systems application. e-Prime – Adv. Electr. Eng., Electron. Energy 10, 100829 (2024).
Article  Google Scholar 
Venkatesh, S. N., Sugumaran, V., Subramanian, B., Josephin, J. S. F. & Varuvel, E. G. A comparative study on bayes classifier for detecting photovoltaic module visual faults using deep learning features. Sustain. Energy Technol. Assess. 64, 103713 (2024).
Google Scholar 
Hong, Y. Y. & Pula, R. A. Diagnosis of photovoltaic faults using digital twin and PSO-optimized shifted window transformer. Appl. Soft Comput. 150, 111092 (2024).
Article  Google Scholar 
Amiri, A. F., Kichou, S., Oudira, H., Chouder, A. & Silvestre, S. Fault Detection and Diagnosis of a Photovoltaic System Based on Deep Learning Using the Combination of a Convolutional Neural Network (CNN) and Bidirectional Gated Recurrent Unit (Bi-GRU). Sustain. (Switzerland) 16, 1012 (2024).
ADS  Google Scholar 
Lu, S. D., Wu, C. C. & Sian, H. W. A novel fault diagnosis method for PV arrays using convolutional extension neural network with symmetrized dot pattern analysis. IET Sci., Meas. Technol. 18, 49–64 (2024).
Article  Google Scholar 
Wang, M. H., Hung, C. C., Lu, S. D., Lin, Z. H. & Kuo, C. C. Fault Diagnosis for PV Modules Based on AlexNet and Symmetrized Dot Pattern. Energies (Basel) 16, 7563 (2023).
Article  Google Scholar 
Chouder, A. & Silvestre, S. Automatic supervision and fault detection of PV systems based on power losses analysis. Energy Convers. Manag. 51, 1929–1937 (2010).
Article  ADS  Google Scholar 
Download references
The authors would like to acknowledge the Deanship of Graduate Studies and Scientific Research, Taif University for funding this work.
This work is funded and supported by the Deanship of Graduate Studies and Scientific Research, Taif University.
Electric Power and Energy Systems Research Laboratory, Ecole Supérieure en Génie Electrique Et Energétique d’Oran, 31000, Oran, Algeria
Abdellatif Seghiour, Yacine Bendjeddou & Imene Meriem Mostefaoui
Electrical Engineering Laboratory (LGE), University of M’sila, BP 166, 28000, M’Sila, Algeria
Aissa Chouder
Department of Electrical Engineering, College of Engineering, Taif University, 21944, Taif, Saudi Arabia
Hisham Alharbi & Abdulrahman Babqi
Department of Electrical Engineering, Umm Al-Qura University, 21421, Makkah, Saudi Arabia
Abdullah S. Bin Humayd
Department of Electrical and Computer Engineering, Faculty of Technology, Debre Markos University, P. BOX 269, Debre Markos, Ethiopia
Abebe Wondiferaw Wondimeneh
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Abdellatif Seghiour, Yacine Bendjeddou, Imene Meriem Mostefaoui, Aissa Chouder: Conceptualization, Methodology, Software, Visualization, Investigation, Writing- Original draft preparation. Hisham Alharbi, Abdullah S. Bin Humayd, Abebe Wondiferaw Wondimeneh, Abdulrahman Babqi: Data curation, Validation, Supervision, Resources, Writing—Review & Editing, Project administration, Funding Acquisition.
Correspondence to Abdellatif Seghiour or Abebe Wondiferaw Wondimeneh.
The authors declare no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Reprints and permissions
Seghiour, A., Bendjeddou, Y., Mostefaoui, I.M. et al. Fault detection and diagnosis in photovoltaic systems using artificial intelligence and time–frequency analysis. Sci Rep 16, 10056 (2026). https://doi.org/10.1038/s41598-026-39386-7
Download citation
Received: 01 June 2025
Accepted: 04 February 2026
Published: 20 February 2026
Version of record: 26 March 2026
DOI: https://doi.org/10.1038/s41598-026-39386-7
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative
Advertisement
Scientific Reports (Sci Rep)
ISSN 2045-2322 (online)
© 2026 Springer Nature Limited
Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

source