The global energy transition is increasingly reliant on renewable sources, with solar photovoltaic (PV) systems playing a pivotal role. A critical component in any PV system is the grid-tied inverter, which serves as the essential interface converting direct current (DC) from solar panels into grid-compliant alternating current (AC). The reliability and efficiency of the entire power generation system are therefore directly tied to the health of the grid-tied inverter. Among the various potential failures within an inverter, open-circuit faults in its power semiconductor switches (IGBTs/MOSFETs) are particularly insidious. Unlike short-circuit faults which typically trigger immediate protective shutdowns, an open-circuit fault allows the system to continue operating, but with degraded performance, increased stress on healthy components, and the risk of secondary failures. This underscores the necessity for robust, intelligent fault diagnosis methods to ensure system safety, maximize availability, and reduce maintenance costs.
This article addresses the challenge of diagnosing single-switch open-circuit faults in three-phase, two-level photovoltaic grid-tied inverters. We propose a novel, two-stage diagnostic methodology that combines signal separation for fault detection with machine learning for precise fault localization. The first stage employs the Fast Independent Component Analysis (FastICA) algorithm to analyze the three-phase output currents and definitively detect the occurrence of a single-switch open-circuit fault. The second stage extracts a comprehensive set of time-domain features from the transformed d-axis current and utilizes a Support Vector Machine (SVM) classifier, optimized with a specific kernel function, to accurately identify which specific switch has failed. The proposed FastICA-SVM framework is validated through detailed MATLAB/Simulink simulations, demonstrating high diagnostic accuracy and robustness under varying power conditions.
The typical architecture of a two-stage PV grid-tied inverter system is shown in the conceptual diagram below. It consists of the PV array, a DC-DC boost converter for Maximum Power Point Tracking (MPPT), the crucial three-phase Voltage Source Inverter (VSI), an LCL filter for harmonic mitigation, and finally the utility grid.
The inverter control strategy typically employs a dual-loop scheme. The outer loop regulates the DC-link voltage to maintain it at the optimal point from the MPPT algorithm. The inner loop controls the dq-axis currents in the synchronous reference frame, with the q-current reference set to zero for unity power factor operation, ensuring efficient power transfer to the grid. The three-phase modulation signals are generated via SPWM or SVPWM techniques.
1. Open-Circuit Fault Analysis in Grid-Tied Inverters
In a three-phase two-level VSI bridge, each leg has two complementary switches (e.g., VT1 & VT4 for phase A). When a single switch, such as VT1, experiences an open-circuit fault, its gate signal becomes ineffective. The consequence is a distinct distortion in the output current waveform of the associated phase. For a phase current flowing out of the inverter (defined as positive), the normal path through the top switch (VT1) is blocked. Consequently, the current is forced to freewheel through the anti-parallel diode of the bottom switch (VD4). This results in the loss of the positive half-cycle of the phase A current. The negative half-cycle remains unaffected as it normally flows through VT4 or VD1. The currents in the two healthy phases (B and C) remain sinusoidal but experience an increase in amplitude to compensate and maintain the power balance. This characteristic waveform—one phase current missing a half-wave and the other two being amplified sinusoids—is the unique signature of a single-switch open-circuit fault in a grid-tied inverter.
2. Proposed FastICA-SVM Diagnostic Methodology
The proposed method is structured into two distinct phases: Fault Detection via FastICA and Fault Localization via SVM.
2.1 Phase 1: Fault Detection using FastICA
Under a single-switch fault, the three-phase currents collectively contain only two fundamental independent waveform shapes: a sinusoidal half-wave and a (amplified) sinusoidal wave. The FastICA algorithm is an efficient blind source separation technique that can recover independent source signals from their observed linear mixtures without prior knowledge of the mixing process. It operates by maximizing the non-Gaussianity (often measured by negentropy) of the estimated components.
Given observed signals \( \mathbf{X} = [i_a, i_b, i_c]^T \), assumed to be a linear mixture of unknown independent source signals \( \mathbf{S} \), the model is:
$$ \mathbf{X} = \mathbf{A} \mathbf{S} $$
where \( \mathbf{A} \) is an unknown mixing matrix. FastICA aims to find a demixing matrix \( \mathbf{W} \approx \mathbf{A}^{-1} \) such that:
$$ \mathbf{Y} = \mathbf{W} \mathbf{X} $$
where \( \mathbf{Y} \) contains the estimated independent components (ICs). When applied to the three-phase currents from a healthy inverter, the dominant ICs will be sinusoidal. When applied to currents from an inverter with a single open-circuit fault, FastICA successfully separates the mixture into its constituent independent waveforms: one IC resembling a sinusoidal half-wave, another resembling a (near) sinusoidal wave, and often a third IC representing noise or residuals. The visual identification of these two specific waveform types in the separated components serves as a definitive, data-driven detector for the presence of a single-switch open-circuit fault in the grid-tied inverter.
2.2 Phase 2: Fault Localization using SVM with Time-Domain Features
While FastICA effectively detects a fault, it does not inherently identify the faulty switch. For localization, we shift analysis to the d-axis current (\(I_d\)) in the rotating reference frame, obtained via the Park transform of the three-phase currents (\(i_a, i_b, i_c\)). The \(I_d\) component, which is directly related to the active power, exhibits distinct patterns for different fault locations, especially under varying irradiation (power) levels.
To characterize these patterns, we extract a comprehensive set of 17 statistical and spectral features from the \(I_d\) signal for each operating condition. This set includes standard time-domain descriptors and frequency-domain moments.
| Feature Category | Feature Name | Mathematical Expression |
|---|---|---|
| Time-Domain | Mean | $$ \bar{x} = \frac{1}{N}\sum_{i=1}^{N} x_i $$ |
| Standard Deviation | $$ \sigma = \sqrt{\frac{1}{N}\sum_{i=1}^{N} (x_i – \bar{x})^2} $$ | |
| Skewness | $$ \frac{\frac{1}{N}\sum_{i=1}^{N} (x_i – \bar{x})^3}{\sigma^3} $$ | |
| Kurtosis | $$ \frac{\frac{1}{N}\sum_{i=1}^{N} (x_i – \bar{x})^4}{\sigma^4} $$ | |
| Maximum | $$ x_{max} $$ | |
| Minimum | $$ x_{min} $$ | |
| Peak-to-Peak | $$ x_{max} – x_{min} $$ | |
| Root Mean Square | $$ RMS = \sqrt{\frac{1}{N}\sum_{i=1}^{N} x_i^2} $$ | |
| Crest Factor | $$ \frac{x_{max}}{RMS} $$ | |
| Shape Factor | $$ \frac{RMS}{\frac{1}{N}\sum_{i=1}^{N} |x_i|} $$ | |
| Impulse Factor | $$ \frac{x_{max}}{\frac{1}{N}\sum_{i=1}^{N} |x_i|} $$ | |
| Clearance Factor | $$ \frac{x_{max}}{(\frac{1}{N}\sum_{i=1}^{N} \sqrt{|x_i|})^2} $$ | |
| Energy | $$ \sum_{i=1}^{N} x_i^2 $$ | |
| Frequency-Domain Moments | Mean Frequency | $$ F_{mean} = \frac{\sum_{k=1}^{M} f_k \cdot S_k}{\sum_{k=1}^{M} S_k} $$ |
| Frequency Standard Deviation | $$ F_{std} = \sqrt{\frac{\sum_{k=1}^{M} (f_k – F_{mean})^2 \cdot S_k}{\sum_{k=1}^{M} S_k}} $$ | |
| Spectrum Centroid | $$ C = \frac{\sum_{k=1}^{M} f_k \cdot S_k}{\sum_{k=1}^{M} S_k} $$ (Often same as F_mean) | |
| Root Mean Square Frequency | $$ F_{rms} = \sqrt{\frac{\sum_{k=1}^{M} f_k^2 \cdot S_k}{\sum_{k=1}^{M} S_k}} $$ |
Here, \(x_i\) represents the sampled \(I_d\) signal, \(N\) is the number of samples, \(f_k\) is the frequency at bin \(k\), \(S_k\) is the spectral magnitude at bin \(k\), and \(M\) is the number of frequency bins.
These 17-dimensional feature vectors, representing different operating states, are used to train a multi-class Support Vector Machine (SVM) classifier. SVM is chosen for its effectiveness in high-dimensional, small-to-medium sample size problems. It finds the optimal hyperplane that maximally separates data points of different classes. For non-linear separation, data is mapped to a higher-dimensional space using a kernel function \(K(\mathbf{x}_i, \mathbf{x}_j)\). The optimization problem is:
$$ \max_{\alpha} \sum_{i=1}^{n} \alpha_i – \frac{1}{2} \sum_{i=1}^{n}\sum_{j=1}^{n} \alpha_i \alpha_j y_i y_j K(\mathbf{x}_i, \mathbf{x}_j) $$
$$ \text{subject to: } \sum_{i=1}^{n} \alpha_i y_i = 0, \quad 0 \leq \alpha_i \leq C $$
where \(\alpha_i\) are Lagrange multipliers, \(C\) is a regularization parameter, \(y_i\) are class labels, and \(\mathbf{x}_i\) are the feature vectors. The classification decision function is:
$$ f(\mathbf{x}) = \text{sgn}\left( \sum_{i=1}^{n} \alpha_i y_i K(\mathbf{x}_i, \mathbf{x}) + b \right) $$
We investigate the performance of SVM with different kernel functions for this fault localization task in a grid-tied inverter:
• Linear Kernel: \( K(\mathbf{x}_i, \mathbf{x}_j) = \mathbf{x}_i^T \mathbf{x}_j \)
• Quadratic Kernel: \( K(\mathbf{x}_i, \mathbf{x}_j) = (\mathbf{x}_i^T \mathbf{x}_j + 1)^2 \)
• Cubic Kernel: \( K(\mathbf{x}_i, \mathbf{x}_j) = (\mathbf{x}_i^T \mathbf{x}_j + 1)^3 \)
• Gaussian (RBF) Kernel: \( K(\mathbf{x}_i, \mathbf{x}_j) = \exp(-\gamma \|\mathbf{x}_i – \mathbf{x}_j\|^2) \)
3. Simulation Results and Analysis
A 100 kW three-phase two-level PV grid-tied inverter system was modeled in MATLAB/Simulink. The system parameters include a DC-link voltage of 800 V, an LCL filter (0.8 mH, 0.2 mH, 100 µF, 3 Ω damping resistor), and a 50 Hz, 380 V (line-to-line) grid. Open-circuit faults were simulated by forcibly setting the gate signal of a specific switch to zero.
3.1 Fault Detection with FastICA: Under a VT1 open-circuit fault, phase A current loses its positive half-cycle, as predicted. Applying FastICA to the three-phase currents post-fault yields three independent components: one clear half-sinusoid, one near-perfect sinusoid, and a residual noise component. This outcome provides unambiguous evidence of a single-switch open-circuit fault, successfully fulfilling the detection phase.
3.2 Dataset Creation for Localization: For fault localization, seven system states were defined: Normal (label 0) and single open-circuit faults for switches VT1 through VT6 (labels 1-6). To ensure robustness, the PV irradiation was varied widely for each state, generating 40 different \(I_d\) current profiles per state, resulting in a total dataset of 280 samples. The 17 features were calculated for each sample, creating a 280×17 feature matrix.
3.3 SVM Model Training and Comparison: The dataset (feature matrix and labels) was used to train and test SVM models with different kernels. Performance was evaluated using k-fold cross-validation (with k varying from 5 to 35). The diagnostic accuracy, defined as the percentage of correctly classified fault states, was the primary metric.
The results, summarized in the table below, show a clear performance trend for the grid-tied inverter fault diagnosis task:
| Cross-Validation Fold | Linear SVM (%) | Quadratic SVM (%) | Cubic SVM (%) | Gaussian SVM (%) |
|---|---|---|---|---|
| 5 | 96.3 | 97.2 | 96.7 | 68.0 |
| 10 | 96.6 | 97.8 | 95.8 | 69.4 |
| 15 | 95.7 | 97.4 | 96.8 | 68.4 |
| 20 | 96.2 | 98.1 | 96.2 | 67.2 |
| 25 | 96.7 | 98.3 | 96.4 | 69.7 |
| 30 | 95.5 | 97.7 | 96.0 | 68.4 |
| 35 | 96.1 | 98.4 | 96.7 | 68.8 |
The Quadratic SVM consistently delivered the highest accuracy, exceeding 97% across all validation scenarios and reaching up to 98.4%. The Linear and Cubic kernels also performed well, with accuracies above 95.5%. Notably, the Gaussian (RBF) kernel performed poorly for this specific feature set and problem, with accuracies below 70%, likely due to overfitting or suboptimal parameter tuning. The superior performance of the quadratic kernel suggests the optimal separating hyperplane in the transformed feature space for this grid-tied inverter problem is best represented by a second-order polynomial.
Further evaluation of the best model (Quadratic SVM with 25% hold-out validation) using the Area Under the ROC Curve (AUC) metric for each class confirmed its high reliability, with AUC values at or very near 1.00 for all fault states and the normal state.
4. Conclusion
This article presented a comprehensive and effective two-stage methodology for the diagnosis of single-switch open-circuit faults in three-phase photovoltaic grid-tied inverters. The integration of FastICA for initial fault detection and an SVM classifier with time-domain features for precise fault localization creates a powerful diagnostic framework. The FastICA stage provides a principled, signal-based method to confirm the occurrence of the specific fault type by isolating its characteristic independent waveforms. The SVM stage, particularly when configured with a quadratic kernel function, proves highly effective in distinguishing between the seven possible operating states (normal + six single-switch faults) with an accuracy consistently above 97%. This high level of accuracy, maintained under varying irradiation conditions, demonstrates the robustness and practical potential of the proposed FastICA-SVM approach for enhancing the reliability and maintainability of solar PV systems. Future work could involve testing the method on hardware-in-the-loop platforms and extending it to diagnose multiple or incipient faults in grid-tied inverters.