In modern photovoltaic systems, various types of solar inverters are deployed to convert DC power into grid-compliant AC power. Among these types of solar inverters, the three-level neutral point clamped (NPC) inverter is widely adopted due to its low harmonic distortion, high efficiency, and even switching loss distribution. However, prolonged operation under harsh conditions often leads to open-circuit faults in the switching devices. Among the different types of solar inverters, the NPC topology is particularly sensitive to faults in the inner switches and switches in the same bridge arm, which are notoriously difficult to diagnose using conventional current or voltage signals. In this work, I propose a novel fault diagnosis method that combines symmetric dot pattern (SDP) transformation with an improved convolutional neural network, named ECA-ConvNeXt, to accurately and efficiently identify 79 types of single- and double-switch open-circuit faults in a grid-connected NPC inverter.
1. Selection of Fault Detection Signals
To overcome the indistinguishability of faults such as SA2 (inner single switch) and SA1SA2 (same-arm double switch) using only output current and line voltage, I define a new set of signals: the three-phase voltages between the upper and lower bridge arm midpoints, denoted as UMNX (X = A, B, C). Together with the A-phase output current iA, these four signals provide sufficient information to discriminate all 79 fault types. The relationship between fault types and labels is summarized in Table A1 (see Appendix). For brevity, only representative examples are listed. The key advantage is that only four sensors are required (one current sensor and three voltage sensors), significantly reducing hardware complexity compared to existing methods that require six or nine sensors.
The average bridge‑arm midpoint voltage for phase A is defined as:
$$ U_{MNA} = V_{A1} – V_{A4} $$
where VA1 is the voltage at the upper midpoint and VA4 is the voltage at the lower midpoint. Under normal operation, UMNA oscillates symmetrically around zero, while during a fault its waveform distorts distinctively.
2. Symmetric Dot Pattern for Multi‑Source Fusion
The Symmetric Dot Pattern (SDP) algorithm maps each time‑domain signal onto a polar coordinate system. For a set of signals, each signal is assigned to a specific sector, and the resulting polar plot resembles a “snowflake” that compactly encodes the relationships among all signals. The transformation rules are:
$$ \begin{aligned}
r(i) &= \frac{x_i – x_{\min}}{x_{\max} – x_{\min}},\\[4pt]
\theta(i) &= \theta_m + \xi \cdot \frac{x_{i+t} – x_{\min}}{x_{\max} – x_{\min}},\\[4pt]
\phi(i) &= \theta_m – \xi \cdot \frac{x_{i+t} – x_{\min}}{x_{\max} – x_{\min}},
\end{aligned} $$
where \(x_i\) is the i‑th sample, \(t\) is the time lag, \(\theta_m\) is the mirror axis angle of the m‑th channel, and \(\xi\) is the gain angle. In my work, four channels correspond to UMNA, UMNB, UMNC and iA, generating a 128×128×1 snowflake image as input to the neural network.
Compared with directly stacking the four signals into a 96×96 matrix, the SDP‑based image provides richer discriminative features. For example, the snowflake patterns for fault types SA2, SA1SA2, SA3, and SA3SA4 are clearly distinguishable from each other (see the experimental results).
3. ECA‑ConvNeXt Model Design
The base ConvNeXt T model is known for merging advantages of Swin Transformer and convolutional networks. However, its high parameter count (27.86 M) and floating‑point operations (1452 M) lead to long inference times, which is undesirable for real‑time fault diagnosis of different types of solar inverters. I therefore propose the ECA‑ConvNeXt model with the following modifications:
- Dual‑size convolution kernels: Each residual block begins with two parallel paths: one with two cascaded 3×3 depthwise convolutions and the other with a single 5×5 depthwise convolution. This expands the receptive field without increasing the kernel size to 7×7.
- Efficient Channel Attention (ECA): After the activation function, a 1D convolution across channels adaptively computes attention weights, emphasizing informative feature maps.
- ReLU activation: Replace the computationally expensive GELU with ReLU to speed up forward propagation.
- Reduced block stacking and channel expansion ratio: The number of residual blocks in the four stages is set to (1, 1, 3, 1) instead of (3, 3, 9, 3), and the channel expansion ratio is reduced from 4 to 2.
The final architecture achieves only 6.10 M parameters and 301.1 M FLOPs, while maintaining high accuracy. Table 1 summarizes the model complexity.
| Model | Parameters (M) | FLOPs (M) | Inference time (ms) |
|---|---|---|---|
| ConvNeXt T | 27.86 | 1452.0 | 8.998 |
| ConvNeXt T + dual kernel | 28.02 | 1465.0 | 10.732 |
| + block reduction | 10.41 | 527.12 | 4.823 |
| + ECA | 10.42 | 528.06 | 6.864 |
| + ReLU | 10.42 | 528.05 | 6.136 |
| + 2× channel expansion (ECA‑ConvNeXt) | 6.10 | 301.10 | 5.617 |
During training, the batch size is 42, the initial learning rate is 3×10⁻⁴ with a decay of 1×10⁻⁵, the dropout rate is 0.04, and the optimizer is AdamW. The training loss and accuracy converge steadily: after 50 epochs, training accuracy reaches 100%, training loss drops to 0.0018, validation accuracy reaches 99.62%, and validation loss is 0.0282.
4. Experimental Validation
I built an experimental platform consisting of a controller (RTUBOX‑204 with TMS320C28346), driving circuit, three‑phase NPC inverter, and high‑precision current/voltage sampling modules. The sampling frequency is 10 kHz, and the fundamental period is 0.02 s. For each fault type, 200 snowflake images of size 128×128 are generated from 2000 consecutive samples. The dataset is split into training, validation, and test sets in a 7:1:2 ratio.
4.1 Feasibility of the Selected Signals
Three signal combinations are compared: (a) UMNX + iA; (b) line voltages + iA; (c) only iA. Using the same ECA‑ConvNeXt model, combination (a) achieves 100% training accuracy, while (b) and (c) reach only 90.57% and 89.05%, respectively. For the 12 confused fault types (same‑arm double and inner single), combination (a) achieves 100% diagnosis accuracy, whereas (b) and (c) give only 50.83% and 48.75%. This confirms the effectiveness of using UMNX and iA together.
4.2 Benefit of SDP Transformation
Comparing SDP‑based images with direct 2D matrix input, the accuracy of ECA‑ConvNeXt improves from 94.24% to 99.53% (a gain of 5.29%). The t‑SNE visualization of the test set features shows that SDP‑processed features form much tighter clusters with clearer boundaries. Table 2 compares the performance of different data input methods.
| Input method | Model | Accuracy (%) |
|---|---|---|
| SDP image | ECA‑ConvNeXt | 99.53 |
| SDP image | ConvNeXt T | 95.72 |
| 2D matrix | ECA‑ConvNeXt | 94.24 |
| 2D matrix | ConvNeXt T | 91.14 |
4.3 Ablation Study of ECA‑ConvNeXt
To isolate the contribution of each modification, I perform an ablation study. The results are shown in Table 3. Starting from the base ConvNeXt T, adding the dual‑size convolution kernel raises accuracy from 95.72% to 97.52% but increases inference time. Reducing block stacking cuts FLOPs by 64% with a slight accuracy drop of 1.29%. Adding ECA brings accuracy back to 99.56%, while replacing GELU with ReLU reduces inference time by 0.728 ms. Finally, reducing the channel expansion ratio to 2 reduces parameters and FLOPs further with minimal accuracy loss (99.53%).
| Model configuration | Params (M) | FLOPs (M) | Inference time (ms) | Accuracy (%) |
|---|---|---|---|---|
| Base ConvNeXt T | 27.88 | 1452.00 | 8.998 | 95.72 |
| + dual kernel | 28.02 | 1465.00 | 10.732 | 97.52 |
| + block reduction | 10.41 | 527.12 | 4.823 | 96.23 |
| + ECA | 10.42 | 528.06 | 6.864 | 99.56 |
| + ReLU | 10.42 | 528.05 | 6.136 | 99.56 |
| + 2× expansion (ECA‑ConvNeXt) | 6.10 | 301.10 | 5.617 | 99.53 |
4.4 Comparison with Other Networks
The proposed SDP + ECA‑ConvNeXt is compared against SDP + ConvNeXt T, SDP + ResNet34, SDP + VGG16, and SDP + MobileNet V1. The results are given in Table 4. Our model achieves the highest test accuracy of 99.53% with an average test time of 8.82 ms (including SDP preprocessing). It outperforms the lightweight MobileNet V1 by 20.99% in accuracy, while requiring only 6.10 M parameters—much fewer than ResNet34 (21.79 M) and VGG16 (134.58 M). The t‑SNE visualization of the output features further confirms that our model separates the 79 fault clusters with the clearest margins.
| Model | Params (M) | FLOPs (M) | Training time per epoch (s) | Test time (ms) | Accuracy (%) |
|---|---|---|---|---|---|
| SDP + ECA‑ConvNeXt | 6.10 | 301.1 | 10.89 | 8.82 | 99.53 |
| SDP + ConvNeXt T | 27.86 | 1452.0 | 21.67 | 11.10 | 95.72 |
| SDP + ResNet34 | 21.79 | 1176 | 11.89 | 11.31 | 97.26 |
| SDP + VGG16 | 134.58 | 5112.0 | 28.53 | 7.86 | 85.41 |
| SDP + MobileNet V1 | 3.29 | 106.7 | 4.25 | 4.49 | 78.54 |
4.5 Robustness Tests
4.5.1 Noise Robustness
I added Gaussian white noise to the test signals at SNR levels of 20 dB, 25 dB, 30 dB, 35 dB, and 40 dB. The model was trained on clean data. Table 5 shows that the proposed method maintains high accuracy even under severe noise. At 20 dB SNR, accuracy is 95.25%, which is 4.17% higher than the next best model (ConvNeXt T, 91.08%). The drop from 40 dB to 20 dB is only 3.73%, indicating strong noise robustness.
| Model | 40 dB | 35 dB | 30 dB | 25 dB | 20 dB |
|---|---|---|---|---|---|
| SDP + ECA‑ConvNeXt | 98.98 | 98.61 | 97.53 | 96.65 | 95.25 |
| SDP + ConvNeXt T | 94.14 | 93.86 | 93.04 | 92.47 | 91.08 |
| SDP + ResNet34 | 96.20 | 95.92 | 95.51 | 93.86 | 92.24 |
| SDP + VGG16 | 83.54 | 82.12 | 81.01 | 78.48 | 74.53 |
| SDP + MobileNet V1 | 74.96 | 73.04 | 72.90 | 69.46 | 64.59 |
4.5.2 Grid Voltage Fluctuation
I simulated voltage deviations of ±2%, ±3%, and ±4% to mimic real grid variations. Three training sets were prepared, each using data from one fluctuation level, and tested on the other two levels. The averaged test accuracy indicates that the model retains high performance (above 92%) even under a ±4% voltage change. The best performance is observed when the training set matches the ±3% fluctuation range (average 96.84%). This demonstrates the generalization capability of the proposed method across different operating conditions.
4.5.3 Irradiance Variation
Using a MATLAB/Simulink model, I varied the irradiance from 1000 W/m² down to 200 W/m² while keeping temperature at 25 °C. The training set included samples under changing irradiance. After 50 epochs, the training accuracy reached 99.83%, and the test accuracy on faults under varying irradiance was 99.15%, with a confusion matrix showing 100% accuracy for 20 randomly selected fault types. This confirms that the method is insensitive to solar irradiance variations—a critical feature for practical types of solar inverters deployed in outdoor photovoltaic systems.
5. Conclusion
I have presented a fault diagnosis method that combines symmetric dot pattern (SDP) transformation with an improved ECA‑ConvNeXt convolutional neural network for grid‑connected NPC inverters. By using the three‑phase upper‑and‑lower bridge midpoint voltages and the A‑phase current as detection signals, the method successfully distinguishes all 79 open‑circuit fault types, including the notoriously difficult same‑arm double‑switch and inner single‑switch faults. The SDP algorithm fuses these four signals into a single snowflake image, reducing data dimensionality while enriching feature representation. The proposed ECA‑ConvNeXt model achieves an excellent balance between accuracy and speed: a test accuracy of 99.53% with an average inference time of 8.82 ms, using only 6.10 M parameters and 301.1 M FLOPs. Extensive robustness tests confirm that the method maintains high accuracy under noise, grid voltage fluctuations, and varying solar irradiance, making it well‑suited for real‑world deployment across different types of solar inverters. The experimental results demonstrate that this framework can serve as a reliable, low‑cost, and real‑time solution for fault diagnosis in modern photovoltaic systems.