Fault Diagnosis of Grid-Connected Solar Inverters Using Symmetric Dot Pattern and Enhanced ConvNeXt

In the context of modern photovoltaic grid-connected systems, the reliability of power electronic converters is paramount. Among various topologies, the three-level neutral point clamped (NPC) inverter has gained widespread adoption due to its superior performance in terms of reduced harmonic distortion, lower switching losses, and improved output waveform quality. However, the high switching frequency and harsh operating environment increase the probability of open-circuit faults in the power switches. Accurate and rapid diagnosis of these faults is essential to maintain system stability and prevent catastrophic failures. This paper presents a novel fault diagnosis framework for grid-connected NPC inverters by integrating the symmetric dot pattern (SDP) technique with an enhanced convolutional neural network model named ECA-ConvNeXt. The proposed method specifically addresses the challenging issue of distinguishing open-circuit faults between the same-arm dual-switch and inner single-switch configurations, which are often misclassified by conventional approaches. By defining a new set of voltage signals called the “upper-and-lower bridge arm midpoint voltages” and combining them with phase-A output current, we effectively differentiate all 79 types of single and double switch open-circuit faults. The SDP algorithm fuses four signals into a single “snowflake” image, enriching feature representation while reducing computational burden. The ECA-ConvNeXt model incorporates dual-scale convolution kernels and an efficient channel attention mechanism to enhance feature extraction and generalization ability, along with optimized activation functions and residual module stacking to balance accuracy and speed. Extensive experiments demonstrate that the proposed method achieves a diagnostic accuracy of 99.53% with an average testing time of 8.82 ms, significantly outperforming existing models. Robustness tests under noise, grid voltage fluctuations, and varying irradiance conditions further validate the practical applicability of our approach. This work provides a reliable and efficient solution for the fault diagnosis of various types of solar inverter in real-world photovoltaic systems.

Introduction

The integration of solar energy into the power grid relies heavily on grid-connected inverters. Among the various types of solar inverter, the three-level NPC topology is preferred for medium- and high-power applications due to its ability to handle high voltages and produce high-quality output. However, the reliability of such inverters is threatened by switch open-circuit faults, which can degrade performance and even lead to system shutdown. Existing fault diagnosis methods can be broadly categorized into model-based, signal-based, and data-driven approaches. Model-based methods, such as adaptive sliding-mode observers and Kalman filters, require accurate system models and are sensitive to parameter variations. Signal-based methods, including current vector and voltage analysis, are limited in the number of fault types they can distinguish and often lack portability. Data-driven methods, particularly those employing deep learning, have demonstrated strong potential due to their adaptability and independence from explicit models.

Despite progress, several challenges remain. First, distinguishing between same-arm dual-switch open-circuit faults (e.g., SA1SA2) and inner single-switch faults (e.g., SA2) is difficult because their output currents and line voltages are nearly identical. Second, many existing data-driven solutions rely on a large number of sensors or multiple input channels, increasing system complexity and computation time. Third, achieving a balance between diagnostic accuracy and speed is critical for real-time applications. To address these issues, we propose a fault diagnosis framework that combines symmetric dot pattern (SDP) with an enhanced ConvNeXt architecture. The key contributions are as follows:

• We define three-phase “upper-and-lower bridge arm midpoint voltages” (U_AMN, U_BMN, U_CMN) and use them together with phase-A output current to accurately distinguish all 79 types of single and double switch open-circuit faults, including the easily confused same-arm dual-switch and inner single-switch types.
• We apply the SDP algorithm to fuse these four signals into a single 2D “snowflake” image, which not only reduces data dimensionality but also amplifies fault-specific features.
• We design an enhanced convolutional neural network (ECA-ConvNeXt) by introducing dual-size convolution kernels, efficient channel attention (ECA) mechanism, and simplified residual blocks. This model achieves high accuracy while maintaining low computational cost.
• Extensive experiments on a real experimental platform verify the effectiveness, robustness, and efficiency of the proposed method under various operating conditions, including noise, grid voltage fluctuations, and irradiance changes.

The remainder of the paper is organized as follows. Section 2 discusses the selection of fault detection signals. Section 3 details the proposed fault diagnosis methodology, including SDP-based information fusion and the ECA-ConvNeXt model. Section 4 presents experimental results and comparisons. Section 5 concludes the paper.

Fault Detection Signal Selection

For a three-level NPC grid-connected inverter, the topology is well known. In practice, simultaneous faults of more than two switches are rare; therefore, we focus on 79 scenarios covering the healthy state, single-switch open-circuit faults, and double-switch open-circuit faults. Our analysis reveals that when faults such as SA2 (inner single-switch fault) and SA1SA2 (same-arm dual-switch fault) occur, the output currents i_A, i_B, i_C and the line voltage u_AB exhibit almost identical waveforms. Similarly, SA3 and SA3SA4 faults are indistinguishable using these signals. This confusion arises because the topology after these faults becomes symmetric in terms of current paths. To resolve this, we introduce a new set of signals: the voltage between the midpoint of the upper bridge arm (M) and the midpoint of the lower bridge arm (N) for each phase. For phase A, this voltage is denoted as u_AMN. The definition is illustrated conceptually in the inverter topology. By analyzing the waveforms, we observe that u_AMN provides distinct signatures for SA2 vs. SA1SA2 and for SA3 vs. SA3SA4 faults. The same applies to phases B and C. Therefore, by combining the three voltages u_AMN, u_BMN, u_CMN with the phase-A output current i_A, we can uniquely identify all 79 fault types. Table 1 lists the complete fault types and their corresponding labels used in this study. Note that the labels are arbitrary but consistent for training.

Table 1: Fault types and labels for 79 scenarios

Fault Type	Label	Fault Type	Label	Fault Type	Label	Fault Type	Label	Fault Type	Label
None	0	SA2SA3	16	SA1SB2	32	SA3SB2	48	SB1SC2	64
SA1	1	SA2SA4	17	SA1SB3	33	SA3SB3	49	SB1SC3	65
SA2	2	SA3SA4	18	SA1SB4	34	SA3SB4	50	SB1SC4	66
SA3	3	SB1SB2	19	SA1SC1	35	SA3SC1	51	SB2SC1	67
SA4	4	SB1SB3	20	SA1SC2	36	SA3SC2	52	SB2SC2	68
SB1	5	SB1SB4	21	SA1SC3	37	SA3SC3	53	SB2SC3	69
SB2	6	SB2SB3	22	SA1SC4	38	SA3SC4	54	SB2SC4	70
SB3	7	SB2SB4	23	SA2SB1	39	SA4SB1	55	SB3SC1	71
SB4	8	SB3SB4	24	SA2SB2	40	SA4SB2	56	SB3SC2	72
SC1	9	SC1SC2	25	SA2SB3	41	SA4SB3	57	SB3SC3	73
SC2	10	SC1SC3	26	SA2SB4	42	SA4SB4	58	SB3SC4	74
SC3	11	SC1SC4	27	SA2SC1	43	SA4SC1	59	SB4SC1	75
SC4	12	SC2SC3	28	SA2SC2	44	SA4SC2	60	SB4SC2	76
SA1SA2	13	SC2SC4	29	SA2SC3	45	SA4SC3	61	SB4SC3	77
SA1SA3	14	SC3SC4	30	SA2SC4	46	SA4SC4	62	SB4SC4	78
SA1SA4	15	SA1SB1	31	SA3SB1	47	SB1SC1	63	—	—

In summary, the selection of u_AMN, u_BMN, u_CMN, and i_A as fault detection signals is validated by the inability of traditional current and line voltage signals to discriminate between the aforementioned critical fault pairs. This signal set reduces the required number of sensors compared to previous work (e.g., nine voltage sensors in literature) while still enabling complete coverage of all 79 fault types relevant to various types of solar inverter.

Fault Diagnosis Method

SDP-Based Information Fusion

Symmetric dot pattern (SDP) is a powerful technique for transforming multi-dimensional time-series signals into a single two-dimensional image. It maps each sample point to a polar coordinate system, where the radius and angles are determined by the normalized values of the signals. For a given set of n signals, each signal occupies a specific angular sector. The transformation is defined as follows:

For a signal x(i) at the i-th sample point, with time offset t, the polar coordinates (r(i), θ(i), φ(i)) are computed as:

$$ r(i) = \frac{x(i) – x_{\text{min}}}{x_{\text{max}} – x_{\text{min}}} $$

$$ \theta(i) = \theta_m + \frac{x(i+t) – x_{\text{min}}}{x_{\text{max}} – x_{\text{min}}} \cdot \xi $$

$$ \phi(i) = \theta_m – \frac{x(i+t) – x_{\text{min}}}{x_{\text{max}} – x_{\text{min}}} \cdot \xi $$

$$ \theta_m = \frac{360^\circ}{n} \times m, \quad m = 0, 1, \ldots, n-1 $$

where x_min and x_max are the minimum and maximum values of the signal, θ_m is the mirror axis angle for the m-th channel, ξ is the gain factor controlling the spread, and n is the total number of signals. In our case, we fuse four signals (u_AMN, u_BMN, u_CMN, i_A) into one image. The resulting “snowflake” pattern varies distinctively across different fault conditions. For example, the patterns for SA2, SA1SA2, SA3, and SA3SA4 faults show clear visual differences that can be effectively learned by a convolutional neural network. Compared to directly stacking the four signals into a 2D matrix, SDP processing significantly amplifies the feature discrepancies between fault classes. Quantitative analysis in the experimental section confirms an accuracy improvement of over 5% when using SDP.

ECA-ConvNeXt Fault Diagnosis Model

The base architecture we adopt is ConvNeXt T, a modern convolutional network inspired by Swin Transformer designs. It consists of four stages with increasing channel dimensions (96, 192, 384, 768) and residual block counts (3,3,9,3). Each residual block contains a depthwise 7×7 convolution, layer normalization, pointwise convolution, and GELU activation. However, the original ConvNeXt T is computationally heavy and may not achieve optimal accuracy for our specific fault diagnosis task. Therefore, we propose several modifications to create the ECA-ConvNeXt model.

1. Dual-Size Convolution Kernels: We replace the initial depthwise convolution in each block with a parallel structure consisting of two paths: one with two consecutive 3×3 depthwise convolutions and another with a single 5×5 depthwise convolution. This design allows the network to capture features at multiple scales. The outputs are concatenated before further processing. Meanwhile, we reduce the original 7×7 depthwise convolution to a 5×5 kernel to maintain the effective receptive field of 7×7 while decreasing parameters.

2. Efficient Channel Attention (ECA): After the pointwise convolution, we insert the ECA module, which performs global average pooling followed by a 1D convolution whose kernel size is adaptively determined based on the channel number:

$$ k = \left\lfloor \frac{\log_2(c)}{2\gamma} + \frac{b}{\gamma} \right\rfloor_{\text{odd}} $$

where c is the number of channels, and γ, b are hyperparameters (set to 1 and 2 respectively). The 1D convolution captures local cross-channel interactions, and a sigmoid function generates channel weights. This mechanism emphasizes important features while suppressing noise.

3. Activation Function Replacement: We replace the GELU activation with the simpler ReLU:

$$ f(x) = \max(0, x) $$

ReLU has lower computational cost and is sufficient for our task.

4. Block Stacking and Channel Expansion Adjustment: To reduce model complexity, we decrease the number of residual blocks per stage to 1,1,3,1, and reduce the channel expansion factor from 4 to 2. These changes drastically cut the parameter count and FLOPs without sacrificing accuracy.

The improved block structure is depicted conceptually: it starts with dual-scale depthwise convolutions (3×3+3×3 and 5×5), followed by layer normalization, then pointwise convolution, then ReLU activation, then ECA, and finally a second pointwise convolution with a residual connection. The overall ECA-ConvNeXt model is lightweight and efficient.

Experimental Verification

Experimental Setup

We built an experimental platform consisting of a controller (RTUBOX-204 with TMS320C28346), driver circuits, a three-phase NPC inverter, and high-precision current/voltage sampling modules. The sampling frequency was 10 kHz with a 4-second acquisition window for each fault condition. For each of the 79 fault types, we collected 200 samples of “snowflake” images (128×128 pixels, single channel). The dataset was split into training/validation/test sets in a 7:1:2 ratio. The neural network was implemented in PyTorch on a system with an NVIDIA RTX 4060 Ti GPU (8GB) and an Intel i5-12490K CPU. Training hyperparameters: batch size 42, 50 epochs, initial learning rate 3×10⁻⁴, weight decay 1×10⁻⁵, dropout 0.04.

Feasibility of Fault Detection Signals

We compared three signal combinations: (a) u_AMN, u_BMN, u_CMN + i_A (proposed), (b) i_A + three line voltages, and (c) only i_A. All signals were processed by SDP and then classified by the same ECA-ConvNeXt model. The training accuracy for combination (a) reached 100%, while (b) and (c) achieved only 90.57% and 89.05%, respectively. For the 12 confusing fault types (same-arm dual-switch vs. inner single-switch), combination (a) achieved 100% diagnosis accuracy, whereas (b) and (c) performed at 50.83% and 48.75%. t-SNE visualization confirmed that features extracted from combination (a) were well-separated, while others overlapped significantly. These results validate that our selected signals effectively resolve the critical fault ambiguity.

Comparison of Data Processing Methods

We compared the proposed SDP-based image input against directly reshaping the four signals into a 96×96 matrix (2D input). Using the same ECA-ConvNeXt model, SDP achieved 99.53% test accuracy, while the 2D input only reached 94.24% – a 5.29% improvement. Similar gains were observed with the base ConvNeXt T model (95.72% vs. 91.14%). The t-SNE plots of the raw features before training revealed that SDP already provided better class separability. This demonstrates that SDP enhances the distinguishability of fault patterns, making the downstream classification task easier.

Ablation Study of ECA-ConvNeXt

We conducted ablation experiments on the SDP dataset to evaluate each proposed modification. The results are summarized in Table 2.

Table 2: Ablation results on ECA-ConvNeXt model

Model Configuration	Params (M)	FLOPs (M)	Inference Time (ms)	Accuracy (%)
Original ConvNeXt T	27.88	1452.0	8.998	95.72
+ Dual-size conv kernels	28.02	1465.0	10.732	97.52
+ Adjust block stacking	10.41	527.12	4.823	96.23
+ ECA	10.42	528.06	6.864	99.56
+ Replace GELU with ReLU	10.42	528.05	6.136	99.56
+ Reduce expansion factor to 2	6.10	301.10	5.617	99.53

Each modification contributed to either accuracy improvement or complexity reduction. For instance, adding dual-size kernels boosted accuracy by 1.80% but increased inference time. Subsequent adjustments to block stacking drastically reduced parameters and FLOPs while only slightly lowering accuracy. The inclusion of ECA recovered accuracy to 99.56%, and replacing GELU with ReLU further reduced inference time without loss. Finally, lowering the expansion factor from 4 to 2 reduced parameters by 4.32 M and FLOPs by 226.95 M with only a 0.03% accuracy drop, yielding the final model with 6.10 M parameters, 301.1 M FLOPs, 5.617 ms inference time, and 99.53% test accuracy.

Training and Testing Performance

After 50 training epochs, the ECA-ConvNeXt model achieved 100% training accuracy, a training loss of 0.0018, 99.62% validation accuracy, and a validation loss of 0.0282. The confusion matrix for 20 randomly selected fault types (including challenging pairs) showed that only one type (label 34) had 95% accuracy; all others were 100%. This indicates excellent classification capability for all 79 types of solar inverter faults.

Comparison with Other Models

We compared ECA-ConvNeXt with several popular deep learning models: ConvNeXt T, ResNet 34, VGG 16, and MobileNet V1, all using the same SDP dataset. Table 3 presents the results.

Table 3: Performance comparison of different models on SDP dataset

Model	Params (M)	FLOPs (M)	Training Time per Epoch (s)	Avg Test Time (ms)	Test Accuracy (%)
ECA-ConvNeXt (ours)	6.10	301.1	10.89	8.82	99.53
ConvNeXt T	27.86	1452.0	21.67	11.10	95.72
ResNet 34	21.79	1176	11.89	11.31	97.26
VGG 16	134.58	5112.0	28.53	7.86	85.41
MobileNet V1	3.29	106.7	4.25	4.49	78.54

Our ECA-ConvNeXt achieved the highest accuracy (99.53%) while maintaining a moderate inference time of 8.82 ms. Compared to ConvNeXt T, it reduced parameters by 78% and FLOPs by 79%, and improved accuracy by 3.81%. Although MobileNet V1 was faster and lighter, its accuracy was 20.99% lower, making it unsuitable for reliable fault diagnosis. t-SNE visualizations of the output features further confirmed that ECA-ConvNeXt produced the most compact and well-separated clusters for all 79 fault types, while other models exhibited varying degrees of feature mixing.

Robustness Tests

Noise Robustness: We added Gaussian white noise to the test signals at signal-to-noise ratios (SNR) of 20 dB, 25 dB, 30 dB, 35 dB, and 40 dB. The model was trained on clean data. Then we tested the noisy data. The ECA-ConvNeXt consistently outperformed all competitors. For example, at 20 dB SNR, it achieved 95.25% accuracy, while the next best (ResNet 34) was 92.24%. At 40 dB, accuracy reached 98.98%. The accuracy degradation from 40 dB to 20 dB was only 3.73%, demonstrating strong noise immunity.

Grid Voltage Fluctuation Test: We simulated grid voltage variations of ±2%, ±3%, and ±4%. Models were trained on data from one fluctuation range and tested on the others. The average accuracy for training on ±3% and testing on ±2% and ±4% was 96.84%; the worst case (training on ±2%) averaged 94.39%. This indicates that our method maintains high performance even when the grid voltage deviates from the training condition.

Irradiance Variation Test: Using a MATLAB/Simulink simulation, we varied the solar irradiance from 1000 W/m² down to lower values (starting at 0.2 s). The model trained on data including irradiance changes achieved 99.83% training accuracy and 99.15% test accuracy. For 20 randomly selected fault types, the test accuracy was 100% each, proving that the method is robust to changing environmental conditions affecting the photovoltaic source.

Conclusion

In this work, we have developed a high-performance fault diagnosis framework for grid-connected three-level NPC inverters, which are one of the prevalent types of solar inverter in photovoltaic systems. By introducing the novel concept of “upper-and-lower bridge arm midpoint voltages” combined with phase-A current, we successfully resolve the long-standing problem of distinguishing same-arm dual-switch open-circuit faults from inner single-switch faults. The use of symmetric dot pattern (SDP) to fuse four signals into a single image not only reduces data dimensionality but also amplifies discriminative features, leading to a 5.29% accuracy improvement over direct 2D input. The proposed ECA-ConvNeXt model, with its dual-size convolution kernels, efficient channel attention, simplified activation, and optimized architectural parameters, achieves an outstanding 99.53% diagnostic accuracy while maintaining a low inference time of 8.82 ms. Extensive comparative experiments and robustness tests under noise, grid voltage fluctuations, and irradiance variations confirm that our method is both accurate and reliable for real-world applications. The framework offers an excellent balance between precision and computational efficiency, making it suitable for deployment in embedded systems for online fault monitoring of various types of solar inverter. Future work will focus on extending the method to handle more complex multi-switch faults and adapting it to other power converter topologies.