Data-driven health monitoring of power converters remains limited by poor generalization to unseen operating conditions. This work addresses this out-of-distribution (OOD) challenge by building a domain-specific time-series foundation model (PE-TSFM) that learns representations directly from large-scale unlabeled converter data. Unlike generic TSFMs trained on broad time-series datasets, the proposed PE-TSFM is pretrained entirely on domain data, enabling it to learn the physical relationships unique to power electronics. To further tailor the model to this domain, we introduce a dual-attention mechanism that captures both temporal patterns and inter-channel dependencies. While generic TSFMs primarily model temporal dependencies, the added channel attention (CA) captures inter-sensor physical relationships essential for converter degradation analysis. A dataset containing 141 million unlabeled timestamps from an operating power converter is used for pretraining. Experiments show that PE-TSFMs achieve 76%-92% accuracy under unseen operating conditions. In contrast, generic TSFMs achieve around 60% and conventional time-series models achieve around 40% accuracy. This result confirms the strong OOD generalization of the proposed PE-TSFM. Ablation studies further verify that the introduced CA mechanism significantly improves model performance. In addition, we conduct detailed studies on model scalability, hyperparameter sensitivity, and interpretability to provide a comprehensive understanding of the proposed approach.

Accurate prediction of operation rhythm and blast furnace gas (BFG) consumption for hot blast stoves group are critical for energy conservation and consumption reduction in steel industry. Considering the spatio-temporal non-stationarity, the multi-modal, and the strong nonlinearity characteristics caused by the alternating uncertain operation of multiple stoves, this paper proposes a process-constrained spatio-temporal graph neural network (PC-STGNN) prediction model. By integrating process mechanisms, the model extracts the hot blast stove operation rhythm, the remaining state time, and the flow rate change treating as process-constrained features, i.e., the industrial expert knowledge embedded in the data input layer. In terms of architectural design, a serial spatio-temporal fusion feature extraction mechanism is constructed, which utilizes the graph attention network (GAT) to dynamically capture the spatial nonlinear coupling weights that evolve with inter-stove conditions. To address the sharp fluctuations in energy consumption patterns, a mixture of experts (MoE)-based hierarchical prediction strategy is introduced to achieve adaptive regression of consumption patterns under different modes such as on-blast, combustion, and stove-switching. Real-world data from the SCADA system of a large-scale steel enterprise in China are employed in the experiments. It demonstrates that the PC-STGNN model significantly outperforms traditional time series models. The proposed method effectively improves the prediction performance when the sudden change occurs.

Failure mechanism analysis of IGBT modules is essential for the reliability assessment in AC converters. Standard DC power cycling tests (PCT) fail to fully replicate the stress conditions encountered in actual AC converters. To address this issue, this article investigates failure mechanisms in IGBT modules under AC repetitive stresses at the converter level and compares the failure difference under DC and AC PCTs. Under the same junction-temperature fluctuation and maximum junction temperature, experiments reveal that DC PCTs shorten cycling lifetime with severe bond-wire cracking and solder degradation, while AC PCTs show fewer cracks with metallic spherical particles. An electro-thermal-mechanical multi-physics model incorporating creep/viscoplasticity clarifies that longer high-temperature dwell and higher temperature change in DC PCTs accelerate strain accumulation, driving faster degradation. Consequently, in a practical converter, IGBT degradation under AC repetitive stresses should be paid more attention to obtain a more realistic degradation trajectory.

Accurate monitoring of junction temperature is essential for ensuring the reliable operation of silicon carbide (SiC) MOSFET-based power converters. This article presents a novel method for estimating junction temperature by utilizing the parasitic inductance voltage during the quasi-turn-on transient. The proposed approach inherently decouples the effects of load current and bypasses high-voltage isolation requirements, thereby eliminating the need for complex multi-parameter calibrations and reducing implementation complexity. The underlying physical mechanism and temperature dependence are examined through an equivalent circuit model, which yields an analytical expression. Double-pulse tests (DPT) verify its performance as a temperature sensitive electrical parameter under various operating conditions. A linearized mapping model that incorporates drain-source voltage facilitates real-time implementation. Experimental verification on a three-phase inverter platform demonstrates excellent agreement with fiber-optic temperature measurements, confirming the method's accuracy, practicality, and suitability for online junction temperature monitoring.

Thermal resistance of the thermal interface materials (TIMs) between power module and heatsink plays a significant role in thermal stress. This article investigates the variances in thermal resistances of TIMs during thermal cycling test and among different chips of a six-switch power module. An experimental setup to facilitate the emulation of practical power state operation is built for the test. The dual interface method with structure function analysis is applied to characterize the TIMs thermal resistances. The results reveal considerable changes and variances in the thermal resistances during the test and among the six switches of the power module, implying high-level of uncertainty caused application-level thermal modeling. A case study of PV inverter application is presented to demonstrate the impact on thermal stress and reliability modeling.

This paper proposes a generic lifetime modeling framework for power semiconductor switches, where the number of cycles to failure is represented as a function of both quantitative stress variables and an explicit technology-conditioning variable. The framework addresses a key limitation of existing lifetime models, which are usually calibrated for specific devices, packaging structures, and test condition ranges. To realize this formulation, a technology-encoded and physics-informed learning approach is developed. The thermal stress variables, including junction temperature swing, maximum junction temperature, and load pulse duration, are transformed into physics-informed features to capture shared degradation trends. In parallel, qualitative device and packaging attributes are encoded into a learnable latent representation. These features are integrated within a dual-path architecture to enable transferable lifetime prediction across devices, packaging configurations, and operating regimes. The proposed framework is trained and evaluated using 578 power-cycling test results collected from the literature over the past 26 years. Experimental validation is further performed on 45 devices under 9 test conditions. The results show that, by leveraging prior lifetime knowledge, a stable device-specific lifetime model can be generated using only 10 new test samples.

The insulated gate bipolar transistor module's degradation monitoring and identification play a crucial role in enhancing reliability. The bond-wire fatigue and solder layer degradation, as two prevalent degradation modes, are interactively coupled, which presents a significant challenge in accurately monitoring and characterizing the failure modes. To address this, a decoupled degradation monitoring method based on temperature-quantified indicators is proposed in this article. A rigorous electrothermal analysis derives the relationship between the temperature difference, power loss, and thermal parameters. The temperature differences at the half-cycle and full-cycle nodes of the power loss period are introduced to construct quantified indicators, which separately characterize two degradation modes. The proposed method only requires the two temperature difference nodes to monitor and quantify the aging states of the bonding wires and the solder layer. This method significantly reduces monitoring complexity and system costs, allowing for direct integration with pre-existing temperature monitoring methods. The simulation and experimental results have verified the validity of the proposed method.

Lifetime prediction is an important but still challenging area of research in the reliability analysis of power semiconductor modules. Considering the nonlinear degradation process and the difficulty in quantifying the analytical expression, the data-driven approaches can be a potential pathway to solve this complex problem. Distinct from the common black-box approaches, the proposed model in this study is constructed by integrating the physical understanding of degradation mechanics through the partial differential equation (PDE) form. In its structure, two additional neural networks are introduced to extract the potential latent variables and explore the degradation dynamics, respectively. Moreover, the time derivative of the lifetime serves as the constraint. Different from the conventional strategy that predicts the future trajectory of the health indicator and further infers the lifetime, the proposed method directly uses the lifetime as the output. Power cycling tests on 18 devices under three test conditions demonstrate the performance of the method and compare it with the trajectory-based data-driven prediction method.

Climate change and grid decarbonisation can alter the energy and carbon profiles of buildings over their life-cycle. Rising average temperatures shift heating and cooling demands, while grid decarbonisation affects the carbon intensity of operational energy. These factors necessitate an adaptive life-cycle analysis approach that accounts for temporal dynamics, ensuring both operational and embodied carbon impacts are addressed with consideration of long-term trends. This study evaluates refurbishment strategies for three UK higher education buildings under varying decarbonisation pathways, using a multi-objective genetic algorithm (NSGA-II) to identify life-cycle carbon and life-cycle cost optimised solutions. Findings reveal that operational efficiency and fuel source significantly shape the life-cycle carbon footprint (LCCF), with electricity-based systems such as air-source heat pumps (ASHPs) achieving notable reductions, even under slower decarbonisation scenarios. The results demonstrate the growing significance of embodied carbon as grid decarbonisation reduces the relative impact of operational emissions, with the embodied component contributing up to 60% of total LCCF in refurbished buildings. Delayed transitions away from gas-based heating prolong emissions growth, whereas ASHP-based solutions under optimistic decarbonisation pathways show potential for long-term carbon stabilisation by the 2040s. Furthermore, grid decarbonisation trajectories influence additional refurbishment priorities, with faster decarbonisation pathways facilitating more nuanced interactions between energy saving measures, and thus greater design flexibility.

Active thermal control (ATC) methods are beneficial to improve the reliability of interactive power converters. However, implementing the ATC methods is associated with several concerns, e.g., voltage fluctuations and current distortions, which degrade the normal operations of interactive power converters. Accordingly, an ATC method based on zero-sequence voltage injection (ZSVI) is proposed for a three-phase pulsewidth modulation (PWM) rectifier in this article. The implementations of the proposed ATC method are detailed, and a thorough performance analysis of the three-phase PWM rectifier with the proposed ATC method is conducted. It is revealed from the analysis that using the proposed method, the DC-link voltage fluctuations and the grid current distortions remain, and then degrade the performance of the three-phase PWM rectifier. Consequently, viable solutions are developed to ensure the normal operation of the three-phase PWM rectifier. Experimental tests are performed to make the verifications for the effectiveness of the proposed method under different cases.

Modular multilevel converters (MMCs) are susceptible to the open-circuit faults of power switches in submodules (SMs). During the fault location interval, the MMC is subject to severe performance degradations, including overcurrent, power fluctuations, and current distortion. These threaten the operating quality and reliability of the MMC, especially under medium-voltage scenarios with stringent operating demands. As a solution, this article proposes a “performance-oriented” seamless fault-riding-through (FRT) strategy, which ensures that the MMC performs as effectively as a fault-free system during the fault location. The seamless FRT operation is achieved by the coordination of the proposed model adaptive predictive control (MAPC) and simultaneous fault location. In the proposed MAPC, the correction factor is estimated to adaptively acquire the precise predictive model of MMC despite the fault occurrence to minimize the tracking errors of control targets. Moreover, the fault location can be achieved with reduced time, even with existing methods. The experiments are conducted to validate the effectiveness of the proposed seamless FRT strategy, which significantly enhances the transient MMC performance quality and reliability for both single and multiple faults.

The increasing power losses in renewable energy converters can accelerate the degradation of power modules and reduce system reliability. This article proposes an enhanced modulation method to address the unbalanced loss distribution in typical renewable converter modules. The approach periodically adjusts switching operation modes to equalize power losses and average the junction temperature of power devices. A full-bridge inverter module is used to demonstrate the method, and experiments confirm that it achieves a more balanced loss distribution compared with traditional modulation. Reliability analysis based on Weibull function-derived time-to-failure distributions and annual mission profiles of a 5 kW photovoltaic (PV) system in Arizona shows that the proposed method can significantly extend the lifetime of power modules without increasing system cost. Furthermore, life cycle assessment (LCA) shows that the proposed method prevents at least two module replacements over a 25-year service life, reducing environmental impact. By improving module longevity, the method also contributes to the overall sustainability of renewable converters.

The electrification of transportation pushes traction inverters, on-board chargers, and aircraft propulsion systems toward higher power density, increasingly enabled by wide-bandgap devices, thereby concentrating ever-larger heat fluxes on power semiconductor dies as package-level cooling nears its sustainable limit. Junction temperature is the resulting bottleneck: it caps the usable rating and, through the thermal cycling that fatigues module interconnects, governs reliability and service life. Managing junction temperature therefore requires a coupled chain of processes spanning the device loss that generates the heat, the models that predict it, the parameters which these models require as input, the cooling that removes it, and the control that bounds its excursion. This review traces heat transfer along this chain, organizing each link by method class and comparing representative methods against the limitation that marks its open problem. The methods within each link are relatively mature, whereas the couplings between links and their validation against realistic missions are not. The cross-cutting gaps, among them sparse drive-cycle validation, weak coupling between electro-thermal and aging models, and the absent co-design of cooling and control, are consolidated into a forward research agenda.

Power cycling tests are widely used to obtain cycle-to-failure data for qualifying the package reliability of power semiconductor modules. In conventional practice, each test must be run to failure, which makes lifetime characterization costly and time-consuming. This paper reveals that early-stage degradation already contains a compact and physically meaningful indicator of device lifetime. An early degradation indicator is identified from the resistance evolution during the initial wear-out stage. Based on this indicator, a data-driven approach is developed to infer cycle-to-failure using only the first 5% of power cycling data. Experimental validation is conducted on 65 power semiconductor modules tested under 11 distinct conditions. The results show a linear relationship, in logarithmic form, between the early-stage degradation rate and cycle-to-failure across all tested conditions. These findings demonstrate that early degradation rates can serve as a reliable lifetime indicator, enabling substantial reduction of test duration without relying on full degradation trajectory prediction.

Bond wire aging monitoring is essential for assessing the health status of power devices in power converters. However, existing approaches typically require electrical connection to the converter and are susceptible to variations in the load current. To address these issues, this paper proposes a non-invasive bond wire aging monitoring method based on the combined magnetic flux density (MFD). First, the MFD characteristics generated by multiple bond wires are thoroughly analyzed, and a new health indicator extracted from the combined MFD is developed, which is sensitive to different aging modes while being independent of temperature and load current. Furthermore, by integrating the minimum inter-class distance (MICD) and between-class variance (BCV), an evaluation system is constructed to determine the optimal sensor position with a minimal number of sensors. Finally, the feasibility of the proposed method is validated through comprehensive investigations, including theoretical analysis, finite-element simulations, and experimental verifications.

A comprehensive review on the development of thermal network models for power semiconductor modules is conducted in this paper. Driven by the wide-spread applications of the power electronics in numerous applications, the development of power modules is revolutionised with higher requirements in power density, switching frequency, operational temperature and reliability, which in turn induces considerable challenges on the thermal management and modelling. As one of the most promising thermal modelling technologies, the thermal networks describe superior performance in long-term profile-based temperature estimation, excellentmulti-physics analysis capability, good hardware compatibility as well as reasonable balance between computational load and accuracy. After revisiting the theoretical basis of the thermal networks, this paper focuses on the evolvement of the thermal networks in terms of format, modelling methodology, thermal boundary condition treatment and verification methods. The state-of-the-art topologies of the representative thermal networks are compared in detail, with a chronology of the thermal networks development being summarised. In addition, the typical application scenarios of the thermal networks and their advantages compared with other technologies are summarised, accompanied by a number of engineering implementation examples. What's more, the future development opportunities and challenges of the thermal networks are discussed, making this paper an all-around reference for researchers and engineers in the power module thermal modelling.

Power cycling tests are essential for reliability qualification of power semiconductor devices, but the required test effort remains costly and time-consuming. Cycle life prediction offers a promising way to reduce the number of fully tested samples. However, when predicted values are further incorporated into lifetime modeling, inaccurate predictions may introduce non-negligible bias into the resulting model, making data screening necessary before the tested and predicted samples are jointly used. This paper proposes a screening-based framework to assess the suitability of predicted cycle life for subsequent lifetime modeling. In the proposed method, samples that are inconsistent with the overall data distribution are identified by an adjusted boxplot, which is well-suited for the skewed distribution of cycle life data (i.e., common Weibull distribution in reliability testing), and are excluded from lifetime model development. The framework is validated using a power cycling dataset of 30 samples tested under five stress conditions. In addition, the relationship between individual prediction error and lifetime model error is systematically analyzed. The results show that the proposed screening method effectively reduces the error of the resulting lifetime model, while the samples identified as outliers are generally associated with larger prediction errors. The study further indicates that the benefit of improving prediction accuracy is more significant when fewer tested samples are available. These findings provide practical guidance for integrating cycle life prediction into reliability-oriented lifetime modeling and for reducing the overall test effort in power cycling qualification.

As the intelligent core of new power systems,electric power intelligent computing systems(EPICS)are undergoing paradigm transformation from information infrastructure to cognitive decision-making enablers.This paper establishes a three-dimensional analytical framework encompassing"theoretical architecture-technological ecosystem-application paradigms"to systematically elucidate the evolutionary logic and development pathways of EPICS.First,we propose a domain-adaptive definition framework that reveals EPICS' core characteristics of"high computing power,high energy efficiency,strong intelligence,and enhanced security",while constructing a hierarchical architecture system comprising perception layer,network layer,platform layer,and application layer.The study further reviews technological advancements through four critical dimensions:heterogeneous computing architectures,intelligent algorithm acceleration,multi-modal data governance,and high-reliability system technologies.Empirical analyses of typical engineering applications demonstrate current implementation achievements and technical bottlenecks.Finally,challenges and future directions are proposed across three levels:theoretical breakthroughs in autonomous decision-making mechanisms,technological innovations in power-computing convergence,and strategic development of standardized evaluation systems.This research provides foundational theoretical references and methodological paradigms for constructing independently controllable electric power intelligent computing systems.

Reliability is a critical performance metric for power semiconductor switches and power electronic systems. Yet guidance on how to test and quantify that reliability is fragmented in the existing literature, particularly with the rapid adoption of wide-bandgap (WBG) devices and novel packaging technologies. This review brings guidance on what designers, reliability engineers, and researchers need to know about power cycling testing (PCT). We provide three major contents: first, introducing how new materials and packaging shift dominant failure mechanisms; second, comparing the main PCT standards joint electron device engineering council (JEDEC), automotive electronics council (AEC), international electrotechnical commission (IEC), and automotive qualification guideline (AQG) and explaining why the "test-to-fail" standard principle is overtaking legacy "test-to-pass" rules; and third, summarizing the unique challenges and existing solutions of applying PCT methods to WBG and ultra-WBG devices. Notably, to the best of the authors' knowledge, this is the first in-depth analysis of the newly released IEC 60749-34:2025 and AQG 324:2025, benchmarked against their earlier editions. Moreover, by collecting more than 200 testing samples from the existing literature, we also offer the first generic lifetime model that spans Si, SiC, multiple bond-wire materials, and die-attach technologies. Finally, the limitations and associated open questions are discussed to identify future research opportunities.