As the global energy landscape shifts toward renewable sources, solar photovoltaic (PV) power generation has emerged as a cornerstone technology. Statistics indicate that the growth of PV installations in China has far exceeded expectations, and the market for PV applications is expected to remain robust for the foreseeable future. However, solar inverters, which serve as the critical interface between the PV array and the electrical grid, face significant reliability challenges due to their complex internal circuitry and harsh operating environments. Among the various components within solar inverters, the insulated gate bipolar transistor (IGBT) is particularly susceptible to faults. Open-circuit faults in IGBTs are especially insidious because they may not immediately shut down the system but can lead to degraded power quality and secondary failures if left undetected. This thesis focuses on developing advanced fault diagnosis methods for IGBT open-circuit faults in two-level solar inverters, leveraging optimized convolutional neural networks (CNNs) combined with time-series modeling and attention mechanisms.
1. Introduction and Research Motivation
The rapid expansion of solar PV systems worldwide has brought increased attention to operational safety and maintenance efficiency. Solar inverters are responsible for converting the DC power generated by PV arrays into AC power suitable for grid integration. Unlike traditional power generation equipment, solar inverters must handle variable input from PV panels, grid disturbances, and environmental stressors such as temperature fluctuations and irradiance changes. These factors make solar inverters one of the most failure-prone components in PV systems. According to a survey of component failure rates, power switching devices account for the largest proportion of failures in solar inverters. Among these, IGBTs are widely used due to their excellent switching characteristics and high efficiency. However, IGBT failures can manifest as either short-circuit or open-circuit faults. Short-circuit faults are typically mitigated by external protection circuits such as fuses, while open-circuit faults are more challenging to detect because they do not immediately trigger protection mechanisms. An open-circuit fault in one or more IGBTs causes current waveform distortion, leading to increased harmonic content, reduced output power, and potential damage to other system components. Therefore, reliable and accurate fault diagnosis for solar inverters is of paramount importance.
Traditional fault diagnosis methods for solar inverters can be categorized into three main groups: signal processing-based, model-based, and data-driven approaches. Signal processing methods often rely on thresholding of specific electrical parameters such as currents, voltages, or power. For instance, wavelet analysis, empirical mode decomposition (EMD), and time-frequency domain analysis have been used to extract fault features. While these methods are straightforward, they require substantial domain expertise and may struggle with complex fault scenarios. Model-based methods use observers or filters, such as Luenberger observers or Kalman filters, to generate residuals that indicate faults. These approaches require accurate mathematical models of the system, which are difficult to obtain for nonlinear and time-varying solar inverters. In recent years, data-driven methods based on deep learning have gained significant traction due to their ability to automatically learn hierarchical feature representations from raw data. Convolutional neural networks (CNNs) have shown exceptional performance in feature extraction, while long short-term memory (LSTM) networks excel at modeling sequential dependencies. By combining these architectures, it is possible to develop robust fault diagnosis systems that can handle the inherent complexity of solar inverter signals.
The work presented in this thesis addresses two key challenges in solar inverter fault diagnosis. First, raw current data from the DC side of solar inverters is often high-dimensional and noisy, making direct classification difficult. To tackle this, a data preprocessing pipeline based on EMD and principal component analysis (PCA) is introduced. Second, because fault signals are essentially time series with temporal correlations, a hybrid CNN-LSTM model is proposed to capture both spatial and temporal features. Furthermore, to improve robustness against environmental noise, an attention mechanism is integrated into the network, allowing the model to focus on the most informative parts of the input sequence. Finally, Bayesian optimization is employed to automatically tune hyperparameters, ensuring near-optimal model performance without exhaustive manual search. The experimental results demonstrate that the proposed methods achieve high diagnostic accuracy and strong robustness under various noise conditions, making them highly suitable for real-world solar inverter monitoring systems.
2. Modeling of PV Power Generation System and Data Acquisition
To generate realistic fault data for solar inverters, a comprehensive simulation model of a PV power generation system was developed in MATLAB/Simulink. The system consists of a PV array, a boost converter, a two-level three-phase inverter, an AC load, and control circuits including maximum power point tracking (MPPT) and voltage-current dual closed-loop control. Figure 1 shows a typical system configuration, but the actual model is described here in detail. The PV array is modeled using a standard single-diode equivalent circuit, with parameters adjusted to produce a nominal DC voltage of 700 V under standard test conditions (STC). The boost converter raises the PV voltage to a level suitable for the inverter input. The inverter uses six IGBTs (T1 through T6) arranged in a three-phase bridge configuration. The output is connected to a resistive-inductive load through an LC filter. The control system employs a perturbation and observation MPPT algorithm to continuously adjust the duty cycle of the boost converter, ensuring that the PV array operates at its maximum power point despite changes in irradiance and temperature. The voltage-current dual closed-loop control regulates the DC bus voltage and the output currents, maintaining high-quality sinusoidal output waveforms.
To verify the correctness of the simulation model, the system was run for 0.5 seconds with an initial irradiance of 700 W/m² and a temperature of 25°C. At t = 0.2 s, the irradiance was stepped to 1000 W/m² and the temperature to 30°C to simulate realistic environmental variations. The DC bus voltage remained stable at approximately 700 V, with only a brief transient during the irradiance change. The three-phase output voltages and currents remained balanced and sinusoidal, as expected. The total harmonic distortion (THD) of the output voltage was measured using the FFT analysis tool in Simulink. With a fundamental frequency of 50 Hz and a maximum frequency of 1000 Hz, the THD was found to be 0.90%, which is well below the 5% limit specified by international standards. This confirms that the simulation model accurately represents the behavior of a practical solar inverter and can be used for fault simulation.
IGBT open-circuit faults were simulated by forcing the gate signals of the corresponding IGBTs to zero. A total of 22 distinct operating states were considered: one healthy state and 21 fault states. These include six single-switch open-circuit faults (each of T1 to T6), six same-side double-switch faults across different phases (e.g., T1 and T3, T1 and T5, T3 and T5, etc.), three single-phase double-switch faults (T1 and T2, T3 and T4, T5 and T6), and six opposite-side double-switch faults across different phases (e.g., T1 and T4, T1 and T6, T3 and T6, etc.). The fault classification is summarized in Table 1.
| Fault type | Faulty switches |
|---|---|
| Single-switch | T1, T2, T3, T4, T5, T6 |
| Same-side double-switch (different phases) | T1T3, T1T5, T3T5, T2T4, T2T6, T4T6 |
| Single-phase double-switch | T1T2, T3T4, T5T6 |
| Opposite-side double-switch (different phases) | T1T4, T1T6, T3T6, T2T3, T2T5, T4T5 |
The DC-side current was sampled at a frequency of 10 kHz for a duration of 0.2 seconds (from 0.25 s to 0.45 s), resulting in 2000 samples per scenario. For each of the 22 states, multiple fault simulations were performed with different seed values to account for minor variations. In total, a dataset of 44,000 samples was collected, with 2,000 samples per class, and stored in .xlsx format. Each sample is a length-2000 time series representing the DC-side current. This signal was chosen because it reflects the overall health of the inverter and can be easily measured in practice. The waveform patterns for different fault conditions are distinctly different, as shown in the original thesis (not reproduced here due to text limitations). For example, normal operation yields a relatively flat current with a slight increase at 0.2 s due to the irradiance step. A single-switch fault causes the current to become asymmetric and oscillatory, while double-switch faults produce more severe distortions.
3. Fault Diagnosis Using DP-CNN-LSTM
3.1 Data Preprocessing with EMD and PCA
Raw current signals from solar inverters are nonlinear and non-stationary, which complicates direct feature extraction. To address this, empirical mode decomposition (EMD) was applied to each current signal. EMD decomposes a signal into a finite number of intrinsic mode functions (IMFs) and a residual trend. The IMFs represent different oscillatory modes embedded in the signal, from high-frequency to low-frequency components. The decomposition process iteratively extracts the highest-frequency component, known as the first IMF, by subtracting the mean of the upper and lower envelopes of the signal. This process is repeated on the residual until it becomes monotonic or has no more than two extrema. The mathematical formulation of EMD can be expressed as:
For a signal $x(t)$, the first IMF $h_1(t)$ is obtained by repeating the sifting process:
$$ x(t) – m(t) = h_1(t) $$
where $m(t)$ is the mean of the upper and lower envelopes. After satisfying the IMF conditions, the residual is $r_1(t) = x(t) – h_1(t)$. The process continues on $r_1(t)$ to obtain $h_2(t)$, and so on. Finally,
$$ x(t) = \sum_{i=1}^{n} h_i(t) + r_n(t) $$
where $n$ is the number of IMFs and $r_n(t)$ is the final residual.
In the experiments, the number of IMFs varied slightly across different fault types. To maintain consistency, the first 10 IMFs were retained for all signals, as this was the minimum number observed. Each IMF is a time series of length 2000. To reduce dimensionality and remove redundant information, PCA was applied to a matrix formed by the 10 IMFs. PCA transforms the data into a set of orthogonal principal components, ordered by variance. The cumulative variance contribution rate was set to 90%, which resulted in retaining 8 principal components. This preprocessing step, dubbed DP (data preprocessing), effectively compresses the input while preserving the essential fault-related features.
3.2 Architecture of the DP-CNN-LSTM Model
The proposed DP-CNN-LSTM model combines CNNs and LSTMs to exploit both spatial and temporal characteristics of the preprocessed data. The overall structure is illustrated in the following table:
| Layer | Type | Output size |
|---|---|---|
| Input | Sequence of length 8 (after PCA) | 8×1 |
| Conv_1 | Convolution (7×1, 64 filters) | 7×64 |
| Maxpool | Max pooling (7×1) | 7×64 |
| Conv_2 | Convolution (4×1, 64 filters) | 4×64 |
| Batch-normal | Batch normalization | 4×64 |
| ReLU_1 | Activation | 4×64 |
| Conv_3 | Convolution (1×1, 64 filters) | 1×64 |
| ReLU_2 | Activation | 1×64 |
| Flatten | Flatten to vector | 64×1 |
| LSTM | LSTM with 6 hidden units | 6×1 |
| ReLU_3 | Activation | 6×1 |
| FC | Fully connected (22) | 22×1 |
| Softmax | Softmax classifier | 22×1 |
| Output | Classification output | 1 |
In this architecture, the convolution layers extract local patterns from the input features, while the max pooling reduces the spatial dimension. The fully connected layer after flattening converts the extracted features into a fixed-size vector. The LSTM layer then captures temporal dependencies across the sequence. Although the input here is not a classic time series after PCA, the ordering of the principal components still reflects certain inter-component relationships, allowing the LSTM to model higher-order interactions. The final softmax layer outputs a probability distribution over the 22 fault classes. The model was trained using the Adam optimizer with a mini-batch size of 2048, a maximum of 400 epochs, an initial learning rate of 0.01, and a learning rate decay factor of 0.5 applied every 700 epochs. The loss function was categorical cross-entropy, and the dataset was shuffled at each epoch.
3.3 Experimental Results and Comparison
To evaluate the proposed DP-CNN-LSTM model, the collected dataset was split into training and test sets with a ratio of 9:1. Thus, 39,600 samples were used for training and 4,400 for testing. The training loss and test accuracy were monitored. The final test accuracy reached 95.2%, and the training loss decreased to 0.271%. The confusion matrix (not shown here) indicated that six fault classes were classified with 100% accuracy, and the majority of the remaining classes had error rates below 5%. Only a few classes showed moderate confusion, likely due to similar current waveform signatures.
To benchmark the performance, several baseline models were compared: a standard CNN, an EMD-CNN, an improved CNN (Are-CNN), and a deep residual network (DRN). All models were evaluated under identical experimental conditions. The comparison results are summarized in Table 3.
| Model | Test accuracy (%) | Training loss (%) |
|---|---|---|
| CNN | 88.4 | 0.447 |
| EMD-CNN | 90.7 | 0.389 |
| Are-CNN | 92.1 | 0.341 |
| DRN | 93.6 | 0.305 |
| DP-CNN-LSTM (proposed) | 95.2 | 0.271 |
From these results, it is clear that the DP-CNN-LSTM model outperforms all baseline methods. The superior performance can be attributed to the combination of EMD-PCA preprocessing, which denoises and compresses the data, and the hybrid CNN-LSTM architecture that effectively captures both local and global features. The EMD and PCA steps eliminate noise and redundant information, allowing the CNN to focus on discriminative patterns, while the LSTM adds an extra level of temporal reasoning that improves classification accuracy.
To test the robustness of the proposed model in the presence of environmental noise, additive white Gaussian noise with different signal-to-noise ratios (SNR) was added to the test data. The SNR values used were 5 dB, 3 dB, 1 dB, and 0 dB. The SNR definition is:
$$ \mathrm{SNR} = 10 \log_{10} \left( \frac{P_s}{P_n} \right) $$
where $P_s$ is the signal power and $P_n$ is the noise power. The model’s performance was evaluated under each noise level, and the results are presented in Table 4.
| SNR (dB) | Test accuracy (%) | Training loss (%) |
|---|---|---|
| 5 | 95.0 | 0.278 |
| 3 | 94.6 | 0.281 |
| 1 | 93.2 | 0.296 |
| 0 | 92.4 | 0.307 |
Even with severe noise (0 dB), the model maintained an accuracy above 92%, demonstrating strong robustness. This indicates that the DP-CNN-LSTM model is suitable for real-world solar inverter environments where noise is inevitable.
4. Fault Diagnosis with Attention Mechanism and Bayesian Optimization
4.1 Attention Mechanism Selection
Although the DP-CNN-LSTM model performed well, further improvements can be achieved by incorporating attention mechanisms. Attention allows the network to focus on the most relevant parts of the input sequence, thereby improving feature extraction and reducing the influence of irrelevant or noisy segments. Several attention mechanisms exist, including channel attention (SE), spatial attention (SAM), and convolutional block attention module (CBAM). After careful consideration, the Squeeze-and-Excitation (SE) attention module was selected due to its simplicity, low computational overhead, and effectiveness in modeling channel-wise dependencies. SE attention operates as follows:
Squeeze: For an input feature map $X \in \mathbb{R}^{H \times W \times C}$, global average pooling is applied to produce a channel descriptor $U \in \mathbb{R}^{1 \times 1 \times C}$:
$$ U_c = \frac{1}{H \times W} \sum_{i=1}^{H} \sum_{j=1}^{W} X_{c}(i,j) $$
Excitation: The descriptor is passed through two fully connected layers to generate channel weights $V$:
$$ V = \sigma\left( W_2 \, \delta\left( W_1 U \right) \right) $$
where $\delta$ is the ReLU function, $\sigma$ is the sigmoid function, $W_1$ is a dimension-reduction layer, and $W_2$ is a dimension-restoring layer.
Scale: The original feature map is re-scaled by the channel weights:
$$ \tilde{X}_c = V_c \cdot X_c $$
This mechanism enables the network to emphasize informative features while suppressing less useful ones, enhancing the model’s ability to extract fault-related signatures even in the presence of noise.
4.2 Enhanced Model with SE Attention
Building upon the DP-CNN-LSTM architecture, a new model was designed by modifying the internal layers and integrating the SE attention module. The revised architecture is summarized in Table 5. The input to the model is the preprocessed data (8 principal components) from the EMD-PCA pipeline. The model consists of two convolutional layers with batch normalization and ReLU activations, followed by a global average pooling layer to obtain channel-wise statistics. The SE module then generates attention weights, which are applied to the feature map via element-wise multiplication. The resulting weighted feature map is flattened and fed into two LSTM layers with 50 and 20 hidden units, respectively. A fully connected layer with softmax produces the final classification. This model allows the network to dynamically adjust feature importance based on the input sequence, improving its generalization and noise resilience.
| Layer | Type | Output size |
|---|---|---|
| Input | Sequence of length 8 | 8×1 |
| Conv_1 | Convolution (7×1, 64 filters) | 7×64 |
| Batch-norm_1 | Batch normalization | 7×64 |
| ReLU_1 | Activation | 7×64 |
| Conv_2 | Convolution (4×1, 64 filters) | 4×64 |
| Batch-norm_2 | Batch normalization | 4×64 |
| ReLU_2 | Activation | 4×64 |
| Gapool | Global average pooling | 1×64 |
| FC_1 | Fully connected (16) | 1×16 |
| ReLU_3 | Activation | 1×16 |
| FC_2 | Fully connected (64) | 1×64 |
| Sigmoid | Activation | 1×64 |
| Multiplication | Element-wise multiply | 4×64 |
| Flatten | Flatten | 256 |
| LSTM_1 | LSTM (50 units) | 50 |
| LSTM_2 | LSTM (20 units) | 20 |
| ReLU_4 | Activation | 20 |
| FC_3 | Fully connected (22) | 22 |
| Softmax | Softmax | 22 |
| Output | Classification output | 1 |
The model was trained with the same settings as before. The test accuracy increased to 96.8%, and the training loss dropped to 0.258%. This confirms that the SE attention mechanism contributes positively to the overall performance compared to the original DP-CNN-LSTM without attention (95.2% accuracy). The improvement is attributed to the attention module’s ability to recalibrate feature responses, suppressing noise-induced variations.
4.3 Bayesian Optimization for Hyperparameter Tuning
Deep learning models contain numerous hyperparameters that significantly influence performance. Manual tuning is tedious and often suboptimal. To automate this process, Bayesian optimization was employed to find the optimal hyperparameters. Bayesian optimization models the objective function using a Gaussian process and iteratively selects hyperparameter configurations to maximize the expected improvement. In this study, three critical hyperparameters were optimized: the number of LSTM hidden units, the regularization coefficient, and the initial learning rate. The search ranges and results are shown in Table 6.
| Parameter | Initial value | Optimized value | Search range |
|---|---|---|---|
| LSTM hidden units | 70 | 98 | [10, 256] |
| Regularization coefficient | 0.001 | 0.003 | [10⁻¹⁰, 0.1] |
| Initial learning rate | 0.01 | 0.0108 | [10⁻⁵, 0.9] |
With these optimized hyperparameters, the model achieved a test accuracy of 99.7% and a training loss of 0.193%. This dramatic improvement demonstrates the effectiveness of Bayesian optimization in fine-tuning the model for solar inverter fault diagnosis. The combination of SE attention and Bayesian optimization yields a highly accurate and robust diagnostic system.
4.4 Robustness Evaluation under Noisy Conditions
To verify the robustness of the enhanced model, the same noise injection procedure was applied. The test was conducted at SNR levels of 5, 3, 1, and 0 dB. The results in Table 7 show that the model retains high accuracy even under severe noise, with an accuracy of 97.5% at 0 dB and a loss of 0.228%. These values are significantly better than those of the DP-CNN-LSTM model without attention, which had 92.4% accuracy at 0 dB. The attention mechanism helps the model focus on reliable signal features, thereby mitigating noise effects.
| SNR (dB) | Test accuracy (%) | Training loss (%) |
|---|---|---|
| 5 | 98.9 | 0.202 |
| 3 | 97.9 | 0.216 |
| 1 | 97.6 | 0.220 |
| 0 | 97.5 | 0.228 |
Besides accuracy and loss, other evaluation metrics such as precision, recall, and F1-score were computed for each scenario. These metrics are defined as:
$$ \text{Precision} = \frac{TP}{TP+FP} $$
$$ \text{Recall} = \frac{TP}{TP+FN} $$
$$ F_1 = \frac{2 \times \text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}} $$
Table 8 lists the average precision, recall, and F1-score for the final model under different noise levels. All metrics remain above 0.95, confirming the model’s excellent classification capability across all classes.
| SNR (dB) | Precision | Recall | F1-score |
|---|---|---|---|
| 5 | 0.986 | 0.983 | 0.985 |
| 3 | 0.976 | 0.972 | 0.975 |
| 1 | 0.971 | 0.967 | 0.968 |
| 0 | 0.966 | 0.959 | 0.963 |
These results indicate that the proposed attention-based Bayesian-optimized CNN-LSTM model is highly effective for IGBT open-circuit fault diagnosis in solar inverters, delivering state-of-the-art performance with superior robustness.
5. Conclusion and Future Work
This thesis has systematically investigated fault diagnosis methods for IGBT open-circuit faults in solar inverters. A comprehensive simulation model of a PV generation system was developed to generate realistic DC-side current data for 22 operating states. Two novel diagnosis models were proposed. The first model, DP-CNN-LSTM, integrates EMD and PCA preprocessing with a hybrid CNN-LSTM architecture. It achieved 95.2% test accuracy under noise-free conditions and maintained above 92% accuracy under 0 dB noise, outperforming conventional CNN variants. The second model enhanced the architecture with an SE attention mechanism and employed Bayesian optimization for hyperparameter tuning. This advanced model achieved 99.7% accuracy on the test set and remained highly robust under noisy conditions, with precision, recall, and F1-score all above 0.95 even at 0 dB SNR.
The key contributions of this thesis are as follows:
1. A realistic simulation model of solar inverters was constructed and validated through THD analysis, providing a reliable platform for fault data generation.
2. A data preprocessing chain combining EMD and PCA effectively reduces dimensionality and noise while preserving fault-relevant information.
3. The hybrid CNN-LSTM architecture leverages both spatial feature extraction and temporal modeling, demonstrating superior performance over single-network approaches.
4. The integration of SE attention allows the network to adaptively focus on important features, improving robustness in noisy environments.
5. Bayesian optimization automates the hyperparameter search, yielding near-optimal model configurations without exhaustive manual trials.
Future research can extend this work in several directions. First, the current study focuses solely on IGBT open-circuit faults. In practice, solar inverters may suffer from other component failures, such as capacitor degradation or sensor malfunctions. A comprehensive multi-fault diagnostic system that covers all major components would increase practical value. Second, the diagnosis methods are offline; real-time implementation on embedded devices for online monitoring is an important next step. Third, the models are trained on simulated data. Transfer learning techniques could be applied to adapt the models to real-world data collected from operating solar inverters. Finally, the recent advances in transformer-based architectures could be explored to capture even longer-range dependencies and further improve diagnostic accuracy. Nevertheless, the findings of this thesis demonstrate that optimized CNN-based deep learning methods hold great promise for enhancing the reliability and safety of solar inverter systems.