In modern photovoltaic power generation systems, the reliability of power electronic converters is paramount. As a pivotal component, the neutral point clamped (NPC) three-level inverter has been widely adopted due to its superior voltage stress suppression and harmonic optimization capabilities. These inverters represent one of the most critical types of solar inverters used in grid-connected applications. However, the multi-switch topology of NPC inverters makes them susceptible to insulated gate bipolar transistor (IGBT) open-circuit faults. To address this challenge, we propose a novel diagnostic method that integrates S-transform (ST) with a residual network enhanced by a parallel multi-dimensional attention (ResNet-PMDA) mechanism.
The increasing demand for renewable energy has driven the development of various types of solar inverters, among which the NPC three-level topology stands out due to its ability to handle higher voltages and reduce harmonic distortion. Nevertheless, the complexity of these systems introduces new failure modes. IGBT open-circuit faults, if not promptly detected, can lead to significant performance degradation, including voltage distortion, current imbalance, and even cascading failures. Traditional diagnostic methods often rely on manual feature engineering or complex signal processing, which may lack robustness in noisy environments.
Our proposed approach addresses these limitations through a three-step pipeline. First, we extract d-axis voltage and current signals from the inverter’s control system. By parameterizing these signals through Park transformation, we derive a virtual resistance metric that serves as an effective fault feature. This transformation reduces the dimensionality of the original three-phase signals into a single representative quantity, thereby simplifying subsequent analysis while preserving critical fault information. Second, we apply S-transform to convert the one-dimensional virtual resistance signals into two-dimensional time-frequency matrices. Unlike conventional time-frequency methods such as Short-Time Fourier Transform (STFT) or wavelet transform (WT), ST offers superior local time-frequency resolution and maintains a direct phase relationship with the original signal. This property is particularly advantageous for capturing subtle fault characteristics that manifest as transient disturbances or harmonic variations. Finally, we feed these time-frequency images into a deep learning model named ResNet-PMDA. The model leverages residual skip connections to mitigate gradient vanishing in deep networks, while the parallel multi-dimensional attention mechanism simultaneously captures channel, spatial, and coordinate dependencies. This dual-domain enhancement allows the model to focus on fault-sensitive features selectively.
To evaluate the effectiveness of our method, we conduct extensive simulations using MATLAB/Simulink. The simulation model replicates a grid-connected NPC three-level inverter with parameters listed in Table 1.
| Parameter | Value |
|---|---|
| Switching Frequency | 5 kHz |
| DC Link Capacitance | 1.200 mF |
| Grid Voltage | 10 kV |
| Grid Frequency | 50 Hz |
| DC Bus Voltage | Customized |
The dataset comprises 73 fault types, including one healthy state, 12 single-switch faults, and 60 double-switch faults. We collect 100 samples per fault type, resulting in a total of 7300 samples. Each sample contains a 2000-point virtual resistance sequence, which is subsequently transformed into a 224×224 pixel image via ST. We split the dataset into training and testing sets with an 8:2 ratio. White Gaussian noise with signal-to-noise ratios (SNR) of 30 dB, 40 dB, and 50 dB is added to simulate realistic operating conditions. The model is trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 32, which we determined through hyperparameter tuning as shown in Table 2.
| Batch Size | Accuracy (No Noise) | Accuracy (30 dB) |
|---|---|---|
| 8 | 98.70% | 90.41% |
| 16 | 99.25% | 92.26% |
| 32 | 99.95% | 94.93% |
| 64 | 98.70% | 92.81% |
The mathematical foundation of S-transform is given by the following expression. For a signal \(x(t)\), the S-transform is defined as:
$$ S(\tau, f) = \int_{-\infty}^{\infty} x(t) \, \omega(\tau – t, f) \, e^{-j 2 \pi f t} \, dt $$
where the Gaussian window \(\omega(t, f)\) is defined as:
$$ \omega(\tau – t, f) = \frac{|f|}{\sqrt{2\pi}} e^{-\frac{f^2 (\tau – t)^2}{2}} $$
In our system, after Park transformation, the d-axis voltage \(u_d\) and current \(i_d\) are obtained. The virtual resistance \(R_d\) is computed as:
$$ R_d = \frac{u_d}{i_d} $$
This \(R_d\) serves as the input for ST, generating a time-frequency representation that captures both temporal and spectral patterns associated with various fault conditions. For example, a healthy inverter produces a stable virtual resistance waveform, whereas a faulty inverter exhibits significant fluctuations. These differences are clearly visible in the ST images, enabling effective classification by the subsequent deep learning model.
The ResNet-PMDA architecture integrates residual blocks and a parallel multi-dimensional attention mechanism. The residual block can be described by the following operation:
$$ y = \mathcal{F}(x, \{W_i\}) + x $$
where \(x\) is the input, \(\mathcal{F}\) represents the residual mapping composed of convolutional layers and batch normalization, and \(y\) is the output. Batch normalization is performed as:
$$ \hat{x}_i = \frac{x_i – \mu}{\sqrt{\sigma^2 + \epsilon}} $$
$$ y_i = \gamma \hat{x}_i + \beta $$
where \(\mu\) and \(\sigma^2\) are the mini-batch mean and variance, \(\epsilon\) is a small constant for numerical stability, and \(\gamma, \beta\) are learnable parameters. The PMDA module combines channel attention, spatial attention, and coordinate attention to refine feature maps from multiple dimensions. This allows the model to emphasize critical features while suppressing irrelevant ones.
We compare our method with alternative approaches in terms of input signal type and network architecture. Table 3 presents the classification accuracy for different input representations, including raw virtual resistance (Rd), S-transform (ST), synchronous extraction transform (SET), STFT, and continuous wavelet transform (CWT), all fed into the same ResNet-PMDA model.
| Input Type | No Noise | 50 dB | 40 dB | 30 dB |
|---|---|---|---|---|
| Rd | 97.40% | 94.14% | 93.15% | 88.42% |
| SET | 99.14% | 95.62% | 94.03% | 88.39% |
| STFT | 94.79% | 94.11% | 93.01% | 89.45% |
| CWT | 99.89% | 96.84% | 94.73% | 90.79% |
| ST (Proposed) | 100.00% | 97.05% | 96.92% | 94.93% |
The results demonstrate that ST consistently achieves the highest accuracy across all noise levels, particularly excelling in high-noise environments. For instance, at 30 dB SNR, ST outperforms Rd, SET, STFT, and CWT by 6.51%, 6.54%, 5.48%, and 4.14%, respectively. This superiority is attributed to ST’s ability to maintain phase information and provide adaptive time-frequency resolution, which is crucial for capturing non-stationary fault signatures in types of solar inverters.
Furthermore, we evaluate the impact of different network backbones. Table 4 compares the performance of CNN-ResNet, CNN-PMDA, ResNet-18, and our proposed ResNet-PMDA model, using ST input in all cases.
| Network | No Noise | 50 dB | 40 dB | 30 dB | MParams | Test Time (s) |
|---|---|---|---|---|---|---|
| CNN-ResNet | 99.11% | 94.79% | 94.32% | 92.74% | 6.40 | 2.14 |
| CNN-PMDA | 98.84% | 93.70% | 93.15% | 90.83% | 8.50 | 4.31 |
| ResNet-18 | 98.90% | 95.89% | 95.75% | 92.94% | 11.22 | 2.97 |
| ResNet-PMDA (Proposed) | 99.95% | 97.05% | 96.92% | 94.93% | 6.60 | 2.25 |
Our model achieves the highest accuracy across all SNR conditions while maintaining a relatively compact model size (6.60 million parameters) and fast inference time. To investigate the contribution of the attention mechanism, we compare PMDA with other popular attention modules, including Squeeze-and-Excitation (SE), Convolutional Block Attention Module (CBAM), and Coordinate Attention (CA). Results are shown in Table 5.
| Attention Module | No Noise | 50 dB | 40 dB | 30 dB |
|---|---|---|---|---|
| ResNet-SE | 98.08% | 94.93% | 92.12% | 84.32% |
| ResNet-CBAM | 99.45% | 95.82% | 94.86% | 88.49% |
| ResNet-CA | 99.18% | 95.55% | 92.81% | 89.86% |
| ResNet-PMDA (Proposed) | 99.95% | 97.05% | 96.92% | 94.93% |
The PMDA module significantly outperforms other attention mechanisms, especially under severe noise conditions. At 30 dB SNR, PMDA improves accuracy by 10.61%, 6.44%, and 5.07% compared to SE, CBAM, and CA, respectively. This improvement is due to PMDA’s ability to simultaneously model channel, spatial, and coordinate relationships, providing a more comprehensive feature refinement.
The continuous development of types of solar inverters has led to more sophisticated topologies, but also new reliability challenges. Our proposed method effectively addresses these challenges by providing a robust, data-driven approach to fault diagnosis. In practical applications, different types of solar inverters may exhibit unique fault signatures; however, the combination of S-transform and ResNet-PMDA can generalize well across various designs due to its adaptive feature extraction capabilities. The diagnostic framework is not limited to NPC inverters and can be extended to other multilevel converter topologies commonly found in modern types of solar inverters.
In conclusion, this study presents a high-performance fault diagnosis method for NPC three-level inverter IGBT open-circuit faults. By combining Park transformation, S-transform, and a customized deep learning model, we achieve 100% classification accuracy in ideal conditions and 94.93% accuracy under 30 dB noise. The experimental results validate the effectiveness of the proposed approach, demonstrating its strong feature extraction ability and noise robustness. Future work will focus on real-time hardware implementation and further optimization for edge deployment in actual photovoltaic systems. The method contributes to the ongoing effort to enhance the reliability and safety of critical types of solar inverters used in renewable energy applications.