In recent years, with the increasing depletion of non-renewable resources such as oil and natural gas, many countries have embarked on a transition to low-carbon energy. Among renewable, low-carbon energy production, solar photovoltaic technology has shown great potential for small-scale energy generation. Since 2005, the photovoltaic manufacturing industry has grown rapidly, and by 2007, it surpassed Japan to become the world’s largest manufacturer of photovoltaic modules. As of 2017, countries like China, the United States, India, Germany, and Japan have actively deployed solar photovoltaic industries, dominating the photovoltaic market. In the future, solar photovoltaics will continue to play a strategic role in the low-carbon energy development of various nations worldwide. The solar inverter is a core component of photovoltaic power generation systems, and if it fails, it not only affects the stable operation of the system but may also cause immeasurable damage to equipment. Therefore, research on fault diagnosis for solar inverters is crucial in terms of equipment loss, cost reduction, and inverter protection.
Currently, fault diagnosis methods for inverters can be broadly classified into three categories: parameter identification-based methods, physical signal analysis-based methods, and artificial intelligence-based methods. Parameter identification methods involve extracting features from measured current-voltage curves to identify faults. Physical signal analysis methods, such as those using output current signals from three-phase voltage inverters, diagnose faults based on changes in instantaneous frequency and angle characteristics of current vectors. Artificial intelligence methods, including wavelet packet decomposition and principal component analysis, have been applied with classifiers like extreme learning machines or radial basis neural networks for fault diagnosis. However, these methods often require professional physical knowledge and complex feature engineering, lacking strong generalizability. In recent years, with the advancement of artificial intelligence technologies, deep learning techniques have been widely used in inverter fault diagnosis. For instance, multilayer perceptrons have been employed to identify fault types and locations in photovoltaic modules, but due to limitations in feature discriminability, they may not effectively diagnose subtle fault types. To improve feature discrimination, some approaches collect signals from voltage, current, and speed changes, combining neural networks with genetic algorithms for fault diagnosis in induction motor voltage source inverters. Although these methods enhance fault diagnosis rates, they often require multiple sensors, increasing equipment costs.
To reduce the number of sensors and improve feature discrimination, we propose a novel fault diagnosis method based on Empirical Mode Decomposition (EMD) and two-dimensional Convolutional Neural Networks (2D-CNN). This method uses a single sensor to collect DC-side current signals from the solar inverter, applies EMD to extract multiple Intrinsic Mode Functions (IMFs) from the current, and combines them with the original data to form two-dimensional feature maps as input to the 2D-CNN. Then, the 2D-CNN is used for feature extraction, and finally, the model diagnoses and classifies multiple open-circuit fault types. Our approach leverages the adaptive signal processing capabilities of EMD for multi-scale feature extraction and the powerful pattern recognition of CNN, making it robust in noisy environments. In this paper, we detail the methodology, experimental setup, and results, demonstrating the effectiveness and superiority of our EMD-CNN-based method for fault diagnosis in solar inverters.
First, let us analyze the photovoltaic power generation system. The system utilizes solar panels to convert sunlight directly into electrical energy. Under光照 conditions, solar cells generate an electromotive force, and series-parallel connections form a solar array to meet the voltage requirements of the system input. The main circuit topology of the photovoltaic power generation system is shown in the figure below, where the left side consists of a photovoltaic array producing DC input, which is converted into three-phase current and voltage through the inverter. In the circuit, R and L represent resistance and inductance, respectively, C is the capacitor, T1 to T6 are Insulated Gate Bipolar Transistors (IGBTs), and S1 to S6 are switches. The system controls the conduction of IGBTs through a modulation circuit.
Common fault types in solar inverters include IGBT short-circuit faults and IGBT open-circuit faults. IGBT short-circuits can be handled directly on the circuit due to their brief duration, with mature solutions available. However, when one or two IGBTs experience open-circuit faults, the inverter may fail to conduct properly. If more IGBTs are open-circuited simultaneously, the inverter ceases to function, making further study irrelevant. Therefore, our focus is on single or dual IGBT open-circuit faults. As illustrated in the system, when one switch is open, single-device faults occur, totaling six cases (T1 to T6 open individually). When two switches are open, dual-device faults occur, which can be categorized into three types: faults in two power devices of the same phase (e.g., T1 and T4, three cases), faults in two power devices of the same half-bridge (e.g., T1 and T3, six cases), and交叉 faults in two power devices (e.g., T1 and T6, six cases). Including the normal state as one class, the total number of IGBT open-circuit fault types is 22. This comprehensive fault set allows us to develop a robust diagnostic model for solar inverters.
For multi-modal signal decomposition and feature extraction, we employ Empirical Mode Decomposition (EMD), also known as Hilbert-Huang Transform (HHT). EMD is a novel adaptive signal time-frequency processing method that extracts multi-scale features from a single signal, making it particularly suitable for analyzing nonlinear and non-stationary signals. EMD decomposes a complex signal into a finite number of Intrinsic Mode Functions (IMFs) and a residual term (res). Each IMF component contains local characteristic signals of the original signal at different time scales. The basic decomposition process of EMD is as follows: First, obtain the extreme points of the original time series x(t). Then, use cubic spline interpolation to fit the upper envelope x1(t) and lower envelope x2(t), calculating the envelope mean m(t) with the formula:
$$m(t) = \frac{x_1(t) + x_2(t)}{2}$$
Next, compute the new sequence h(t):
$$h(t) = x(t) – m(t)$$
Check if h(t) satisfies the two conditions of an IMF: (1) the number of extrema and zero-crossings must either equal or differ by at most one, and (2) the mean value of the envelopes defined by the local maxima and minima is zero. If not, replace x(t) with h(t) and repeat until h(t) meets the criteria, at which point it is considered an IMF. After obtaining an IMF, remove it from the original signal and repeat the process until the remaining signal is monotonic or constant. Thus, through EMD decomposition, the original signal x(t) is expressed as a linear combination of IMFs from high to low frequency and a residual term r(t):
$$x(t) = \sum_{i=1}^{n} c_i(t) + r(t)$$
where c_i(t) is the i-th IMF component, and r(t) is the residual term that does not contain oscillatory modes but reflects the trend of the signal. The first few high-frequency IMFs contain significant and important feature information from the original signal, which is crucial for fault diagnosis in solar inverters.
For feature learning and classification, we use Convolutional Neural Networks (CNNs), specifically a two-dimensional CNN (2D-CNN) model. CNNs are deep learning models designed for processing grid-like data, such as images or time-series represented as matrices. The CNN architecture typically consists of convolutional layers, pooling layers, and fully connected layers. In the convolutional layer, convolution operations are applied to extract features. The mathematical formula for convolution is:
$$y(i,j) = \sum_{m} \sum_{n} x(i+m, j+n) \cdot w(m,n) + b$$
where X is the input matrix, W is the weight parameter matrix (filter), y(i,j) is the output, * denotes the convolution operation, and b is the bias. In our 2D-CNN, we use multiple filters to capture diverse features from the input data. The pooling layer reduces the dimensionality of the feature maps by aggregating local information, commonly through max-pooling or average-pooling. The max-pooling operation can be described as:
$$y_l = \sigma(\text{down}(y_{l-1}))$$
where down(·) represents the downsampling method, and σ(·) is the activation function. Finally, the fully connected layer integrates the extracted features for classification, with the formula:
$$y_l = \sigma\left(\sum_{i} W_i^{l-1} y_{l-1} + b_l\right)$$
where y_l is the output of layer l, W_i^{l-1} are the weight matrices, b_l is the bias, and σ(·) is the activation function such as ReLU. In our model, we use a Softmax layer at the end to output probabilities for each fault class, enabling multi-class diagnosis for solar inverters.
Our proposed EMD-CNN-based fault diagnosis method for solar inverters involves several steps. First, we simulate IGBT open-circuit faults by controlling switch closures and sample the DC-side current signal using a single sensor. This approach minimizes sensor usage, reducing costs and complexity. Second, we set the number of IMFs and perform EMD decomposition on the current signals to obtain IMF components for each fault condition. The EMD process adaptively extracts multi-scale features, which are essential for distinguishing subtle fault patterns in solar inverters. Third, we combine the IMFs with the original data to form two-dimensional feature vectors, creating a rich dataset that captures both time-domain and frequency-domain characteristics. Fourth, we split the fault dataset into training and testing sets, using the training set to train the 2D-CNN model. The 2D-CNN architecture is designed with three convolutional-pooling blocks, a flattening layer, fully connected layers, and a Softmax output layer. This model learns hierarchical features from the input data, enhancing diagnostic accuracy. Fifth, we input the test set into the trained EMD-CNN model for fault identification and classification. The entire process is automated, reducing the need for manual feature engineering and making it applicable to real-world solar inverter systems.
To validate our method, we conducted experiments using a simulation model of a photovoltaic power generation system. The simulation module includes maximum power point tracking and voltage feedforward control functions, mimicking realistic operating conditions. In the simulation, the sampling frequency is set to 10^4 Hz, with a total run time of 4.1 seconds. The system parameters include a resistance of 0.1 Ω, an inductance of 0.002 H, and a three-phase current frequency of 50 Hz. We discard the first 0.1 seconds of data due to initial transients, collecting data from 0.1 to 4.1 seconds. Additionally, at 0.25 seconds, we change the working conditions by increasing the irradiance from 700 W/m² to 1000 W/m², simulating realistic environmental variations for solar inverters. Data collection involves controlling switches to simulate 22 fault types, including normal operation, with 2,000 samples per fault, resulting in a total of 44,000 samples. After EMD decomposition, we aggregate these into a dataset of 440 samples, each representing a fault instance.
Each fault sample consists of a long one-dimensional time series of 2,000 data points. To manage computational complexity and improve model training, we use a sliding window overlap sampling method to segment the data. Specifically, we divide the long sequence into shorter segments of 1,000 data points each, with an overlap of 800 points (80% overlap rate). This approach generates multiple samples from each original sequence, enhancing the dataset size and capturing temporal dependencies. The number of new samples obtained through overlap sampling is calculated as:
$$\text{Samplenum} = \frac{I_{\text{length}} – W_{\text{length}}}{S_{\text{length}}} + 1$$
where I_length is the original data length (2,000), W_length is the window length (1,000), and S_length is the step length (200). For our settings, this yields 6 samples per original sequence, effectively increasing the dataset for training the EMD-CNN model for solar inverter fault diagnosis.
In model training, parameter settings critically impact performance. We experiment with different detection step lengths and sliding window sizes to optimize accuracy. Under a 5 dB noise condition, we test step lengths from 100 to 300 data points and find that a step length of 200 data points (0.02 seconds, matching the phase current period) yields the best fault diagnosis accuracy of 0.9874. Similarly, we evaluate sliding window sizes of 400, 600, 800, 1,000, and 1,200 data points under various noise levels. The results show that a window size of 1,000 data points achieves the highest accuracy across all noise conditions, as summarized in the table below.
| Sliding Window Size (data points) | Accuracy at -1 dB SNR | Accuracy at 0 dB SNR | Accuracy at 1 dB SNR | Accuracy at 3 dB SNR | Accuracy at 5 dB SNR |
|---|---|---|---|---|---|
| 400 | 0.8521 | 0.8765 | 0.8943 | 0.9210 | 0.9456 |
| 600 | 0.8712 | 0.8923 | 0.9087 | 0.9345 | 0.9567 |
| 800 | 0.8898 | 0.9101 | 0.9254 | 0.9489 | 0.9678 |
| 1000 | 0.9123 | 0.9315 | 0.9432 | 0.9621 | 0.9789 |
| 1200 | 0.9012 | 0.9224 | 0.9376 | 0.9554 | 0.9701 |
To assess the robustness of our method, we add Gaussian noise at different signal-to-noise ratios (SNRs) to the dataset, simulating real-world干扰 conditions in solar inverter systems. The SNR is defined as:
$$\text{SNR} = 10 \log_{10}\left(\frac{P_s}{P_n}\right)$$
where P_s is the signal power and P_n is the noise power. We test SNRs of -1 dB, 0 dB, 1 dB, 3 dB, and 5 dB to evaluate performance under varying noise levels. Our EMD-CNN method demonstrates high accuracy even at low SNRs, highlighting its effectiveness for fault diagnosis in noisy environments.
We compare our EMD-CNN method with three traditional approaches: k-Nearest Neighbors (kNN), Support Vector Machines (SVM), and a standard CNN without EMD. For kNN, we set the hyperparameter n_neighbors to 7. For SVM, we use a linear kernel. For the standard CNN, we reshape the input data to [400, 1] and use a similar architecture but without EMD preprocessing. The batch size is set to 600, and the learning rate is 0.0001 for all neural network-based methods. The fault diagnosis accuracies under different noise levels are presented in the table below, showing that our EMD-CNN method consistently outperforms the others, especially in low-SNR conditions.
| Method | Accuracy at -1 dB SNR | Accuracy at 0 dB SNR | Accuracy at 1 dB SNR | Accuracy at 3 dB SNR | Accuracy at 5 dB SNR |
|---|---|---|---|---|---|
| kNN | 0.6023 | 0.6212 | 0.6452 | 0.8068 | 0.9407 |
| SVM | 0.9293 | 0.9545 | 0.9571 | 0.9697 | 0.9848 |
| CNN | 0.9283 | 0.9417 | 0.9567 | 0.9733 | 0.9783 |
| EMD-CNN | 0.9772 | 0.9829 | 0.9867 | 0.9981 | 0.9981 |
The superior performance of EMD-CNN can be attributed to the multi-scale feature extraction by EMD, which enhances the discriminability of fault signatures in solar inverters, and the deep feature learning by CNN, which automates pattern recognition. For instance, at -1 dB SNR, EMD-CNN achieves an accuracy of 0.9772, significantly higher than 0.9293 for SVM and 0.9283 for CNN. This demonstrates the robustness of our method in challenging noise conditions, making it suitable for practical applications where solar inverters operate in variable environments.
In conclusion, we have proposed a novel fault diagnosis method for solar inverters based on EMD and CNN, utilizing DC-side current signals for open-circuit fault monitoring. Through simulation experiments, we validated the method’s performance, showing high diagnostic accuracy and robustness under different noise levels. The EMD process effectively extracts multi-scale features from current signals, while the CNN model learns hierarchical representations for accurate classification. Compared to traditional methods, our approach reduces sensor usage and manual feature engineering, offering a cost-effective and automated solution for solar inverter maintenance. Future work could explore the integration of CNN with Long Short-Term Memory (LSTM) networks to further enhance feature extraction from time-series data, particularly in high-noise scenarios. Additionally, real-world testing on physical solar inverter systems would help refine the method for industrial deployment. Overall, this research contributes to the advancement of intelligent fault diagnosis technologies for renewable energy systems, supporting the reliable operation of solar inverters in global energy networks.