Yang Liu

The data-driven-based fault diagnostic methods have continued to emerge in power electronics systems in recent years. Among them, though the methods that convert one-dimensional (1-D) electrical signals into two-dimensional (2-D) images for classification have shown promise, they often suffer from information loss or excessive computational burden. This article refines the Gramian angular field (GAF) theory and proposes a lossless and efficient method for transforming time series into images, termed extended Gramian angular field (EGAF). Building upon EGAF, an efficient and robust fault diagnostic method is proposed, with implementation demonstrated on a two-level (2L) three-phase (3P) voltage-source-inverter (VSI). The diagnostic method first extracts the current time series under different faults through Clarke transformation and zero-crossing detection, then converts the windowed time series into 2-D matrices using EGAF to construct the diagnostic dataset, and finally implements fault classification through a lightweight convolutional neural network (CNN). The proposed diagnostic method achieved an accuracy of 98.64%, and the experimental results demonstrate its high performance.

Fault diagnosis is essential to ensure the reliability and continuous operation of power converter systems. Despite significant progress over the years, the applicability of diagnosis methods in transient processes remains limited, hindering their practical application, especially in transportation electrification. To tackle this issue, a diagnosis method for voltage-source inverter (VSI) power switch open-circuit (OC) faults with superior transient adaptability is proposed. The method firstly transforms the three-phase (3P) current into a two-phase orthogonal coordinate system and then analyzes the current vector trajectory under different fault conditions. A fast fault detection method is then developed based on trajectory analysis. Furthermore, the trajectory is converted into a 2-D binary matrix, which is then recognized by a lightweight convolutional neural network (CNN) to locate the fault. Compared to existing methods, the proposed method significantly expands the diagnostic range while maintaining high accuracy, particularly excelling in handling transient processes. Physical experiments have verified the method’s performance across a wide range of operating conditions.

Reliable fault diagnosis of DC-DC converters is critical for ensuring the safety and operational continuity of electric vehicles (EVs) and industrial power systems. In practical applications, early detection of progressive degradation can prevent catastrophic failures and significantly reduce maintenance costs. However, developing robust diagnostic models is hindered by sensor noise, environmental disturbances, and the scarcity of labeled fault data—particularly for subtle degradation scenarios where real-world collection is expensive. To address these issues, this paper proposes a transfer learning framework based on Inception-Residual Neural Network (ResNet) + One-Dimensional Convolutional Neural Network (1D-CNN), aiming to enhance diagnostic performance under small-sample conditions. The proposed network extracts discriminative features from simulated voltage signals (source domain) and transfers this knowledge to real-world experimental data (target domain). To reduce domain discrepancies, a Whitening and Coloring Covariance Alignment (WCA) is employed for global feature alignment, while a novel Classwise WCA strategy further refines the alignment at the category level to preserve fault-specific structures. Additionally, the classifier is fine-tuned using L2 Starting Point (L2-SP) regularization, which constrains parameter shift to prevent overfitting under limited supervision. The method is non-intrusive and practical, operating efficiently using only the output voltage ripple. Experimental results validate the effectiveness of the proposed approach, demonstrating a diagnostic accuracy of 91.6% even with limited samples, significantly outperforming conventional transfer learning methods in cross-domain fault identification tasks.

DC/DC converters play a crucial role in electronic systems by providing the necessary power support and ensuring the proper operation of devices. During operation, the converter is exposed to environmental factors, such as temperature and humidity, leading to component degradation and subsequently posing a threat to system reliability and functionality. This arrticle presents an effective degradation detection method for dc/dc converters using feature extraction techniques. The output voltage ripple is used as the degradation indicator, and its values under healthy and degradation conditions are collected. Based on the time domain, frequency domain, and time-frequency domain feature information of the output voltage ripple, a valid feature selection method is designed to select key features, and the Bayesian optimization-based support vector machine is established to detect degradation using the selected key features. Since no extra hardware circuits or elaborate mathematical models are required, the proposed method has low cost and complexity. Simulation and experimental results show the validity and practicability of the proposed method.

Model predictive control has gained increasing popularity in power electronics in recent years for its simple principle and excellent control performance. However, the traditional model prediction is highly dependent on the mathematical model of the exact system, and the mismatch of system parameters may degrade. Moreover, accurate system modeling remains inherently challenging due to the coexistence of inherent uncertainties and time-varying parametric characteristics. In this regard, a model-free predictive control strategy based on hybrid extended state observer is proposed in this article. It constructs an ultralocal model of the system instead of an accurate physical model to reduce dependence on model parameters. The unmodeled portion of the system is compensated into the unknown terms of the ultralocal model to eliminate steady-state errors and improve the system robustness against disturbances. And the effect of error accumulation in case of model parameter mismatch is reduced. Besides, its stability is analyzed in detail. Finally, a three-phase interleaved buck converter is constructed for simulation and experimental verification. The results show the efficacy of the proposed method with strong robustness.

High-speed ground robots moving on unstructured terrains generate intense high-frequency vibrations, leading to LiDAR scan distortions in Lidar-inertial odometry (LIO). Accurate and efficient undistortion is extremely challenging due to (1) rapid and non-smooth state changes during intense vibrations and (2) unpredictable IMU noise coupled with a limited IMU sampling frequency. To address this issue, this paper introduces post-undistortion uncertainty. First, we model the undistortion errors caused by linear and angular vibrations and assign post-undistortion uncertainty to each point. We then leverage this uncertainty to guide point-to-map matching, compute uncertainty-aware residuals, and update the odometry states using an iterated Kalman filter. We conduct vibration-platform and mobile-platform experiments on multiple public datasets as well as our own recordings, demonstrating that our method achieves better performance than other methods when LiDAR undergoes intense vibration.

As critical components in power electronic systems for energy conversion and power management, DC/DC converters are commonly employed in aerospace, electric vehicles, and other electronic devices. Their degradation may lead to instability failure of the power electronic system, significantly impacting the reliability of devices. To prevent irreversible damage, various degradation detection methods have been proposed, which have advantages such as automatic, real-time, accurate, and non-invasive. However, few studies have considered effective degradation detection with reduced resource demands and sampling costs. In this paper, a signal reconstruction technique is adopted in signal acquisition and a novel early degradation detection method is proposed for the output filter capacitor and energy storage inductor of a DC/DC buck converter. In the proposed method, degradation features are extracted from output voltage ripples through long short-term memory fully convolutional network and contrastive learning. Based on the learned features of normal and degraded samples, the converter degradation state is identified. To enhance robustness against noise, the Mixup-based data augmentation is applied to increase data diversity. The effectiveness of the proposed method is verified by simulation and experiment.

With the increasing demand for ocean missions, the development of autonomous marine vehicles (AMVs) has flourished. AMVs rely on electricity to power various systems, so the high reliability and stability of dc-dc converters are critical to the normal operation of AMVs. In reality, fault diagnosis and optimization of converters under the low sampling frequency and computational capability conditions remain practical challenges. This article focuses on the buck converter circuit and utilizes undersampling techniques to obtain output voltage signals sampled at a low sampling frequency. Support vector machine recursive feature elimination is employed for feature selection to reduce computation. Bayesian optimization-based extreme learning machine is used for fault diagnosis, which is suitable for actual deployment and shown to outperform three classical machine learning models. The diagnosis results are used to propose a reliability optimization strategy involving switching frequency adjustment. Physical experiments based on a digital signal processor prove the feasibility of this method.

In electrified transportation systems, power system failures can lead to greater disasters. Therefore, the reliability of converters in transportation systems has been a concern. Fault-tolerant techniques are widely applied to ensure that converters can continue to supply loads under fault conditions. Fault diagnosis as a prerequisite for fault tolerance has also become a research hotspot. This paper proposes a fast method for fault diagnosis of high-frequency LLC converters. The proposed fault diagnosis method is based on the observation of the voltage across the resonant capacitor to determine and locate the faulty power switch, providing a basis for fault tolerance. This diagnosis method requires a voltage sensor, which is also necessary for some control methods. When applying these control methods, the proposed fault diagnosis method can be used without additional sensors, beneficial for cost reduction. A full-bridge LLC converter controlled by a digital signal processor was used as an experimental platform to verify the effectiveness and speed of the proposed diagnostic method. The results show that the proposed fault diagnosis method can achieve the fast diagnosis of high-frequency LLC converters in a short time and with only minimal computational resources.

DC/DC converter is widely used in power electronic applications as a critical part of power supply. During operation, component parameters of the DC/DC converter gradually degrade subjected to various environmental stresses. When they degrade to deviate from the normal value, soft faults will occur and directly affect the functionality and reliability of the whole system. To identify these anomalies effectively, a novel soft fault detection and diagnosis method based on contrastive learning is proposed for a DC/DC Buck converter. First, the output voltage ripples of the converter under healthy and faulty conditions are collected. Then, a feature representation of the voltage ripple is obtained through hierarchical contrastive learning at temporal-level and instance-level. Based on the learned representation, a support vector machine classifier is established to detect and locate soft faults. The proposed method does not need extra hardware or signal injection, thereby reducing the cost and complexity. Simulation and physical experiment results verify the effectiveness and robustness of the proposed method.