Photovoltaic (PV) health monitoring requires fault diagnosis and long-term degradation assessment under variable operating conditions, yet conventional Random Forest (RF) and Long Short-Term Memory (LSTM) models lack physical awareness and rely on single-domain features. This paper proposes a hierarchical physics-enhanced framework that progressively extends RF and LSTM models. A Digital Twin (DT) baseline feeds model-derived residuals to the classifiers; a physics-informed network (PINN-Plain) embeds PV physical relationships and a degradation objective into representation learning; and a Physics-Informed Multimodal Fusion (PINN-MF) framework combines fine-grained electrical measurements, coarse statistical descriptors, and DT residuals through an attention mechanism trained end-to-end. A learned degradation head maps each representation to PV power along a controlled synthetic 25-year ageing scenario. The frameworks are evaluated on a laboratory grid-connected PV emulator covering healthy operation and five fault scenarios under varying irradiance, temperature, and measurement noise. Over 10 independent runs, the RF-path accuracies are 92.78% (DT), 93.38% (PINN-Plain), and 94.14% (PINN-MF), with no statistically significant difference (paired $t$-test, $\alpha$ = 0.05). At 0 dB signal-to-noise ratio, they are 41.97 ± 4.99%, 75.45 ± 5.38%, and 59.24 ± 10.66%, respectively, and the degradation RMSE is 1.79 ± 0.52%, 0.84 ± 0.25%, and 1.23 ± 0.25% (mean ± SD). An ablation, paired $t$-tests, and the seed stability of the learned attention weights, which favour the fine-grained branch in all 10 runs, are reported. These single-testbed results indicate that physics-enhanced multimodal learning is a promising direction for PV fault diagnosis and degradation prediction.
In large-scale solar power plants, photovoltaic (PV) inverters play a central role in energy conversion, which makes understanding their operational reliability essential. The service life of these devices is influenced not only by hardware choices, such as topology and semiconductor technologies, but also by the operational factors embedded in their control, modulation, and supervisory algorithms. This paper presents a comprehensive analysis of how such operational factors drive the reliability of PV inverters. A structured reliability assessment pipeline is employed to estimate device degradation and failure mechanisms under realistic mission profiles. Although several inverter topologies can be adopted in PV systems, this study focuses on how operational procedures influence stress generation, thermal cycling, and lifetime consumption in a 3L-NPC inverter. The operational factors analyzed include current and voltage control strategies, modulation schemes, MPPT algorithms, and grid synchronization techniques. Reliability evaluations are performed using simulations in PLECS and MATLAB/Simulink. The results highlight the operational conditions that most influence inverter reliability. Finally, the study outlines future trends and provides guidelines for Design for Reliability (DfR) in PV inverter operation.
Virtual metrology (VM) can reduce the delay and sampling burden of physical metrology, but its deployment is fragile when sensor layouts, process recipes, or fab conditions change before a sufficiently large target history is available. This paper presents Generative-Few-Shot Virtual Metrology Plus (GFA-VM+), which combines mask-aware automatic sensor alignment, variational representation learning, MAML-based adaptation, and optional FedAvg-style parameter aggregation. The industrial evaluation uses chronological holdouts, paired label draws, and matched label budgets across EPI cross-machine, EPI-to-ETCH cross-process, and early-ramp cross-fab onboarding settings. The results show that the benefit is concentrated in the early calibration regime: ASA prevents structural missingness from being treated as a physical zero, meta initialization accelerates low-label adaptation, and the federated prior contributes most when the target fab has the least local information. At ten target labels, GFA-VM+ achieves mean absolute errors of 1.01 nm for thickness, 3.49 × 1015 cm−3 for doping, 4.51 nm for etch depth, and 3.19 nm for critical dimension; the label efficiency curves enter diminishing-return regimes near nine and eleven labels for cross-machine and cross process transfer, respectively. A controlled synthetic benchmark and sensor-contamination tests further show that the advantage weakens when severe dropout removes shared sensor support or pushes adaptation outside the local optimization regime. Model-update inversion experiments further show that raw-data non-centralization does not eliminate information exposure: batching, clipping, and protected aggregate-only visibility reduce, but do not remove, empirical recoverability. FedAvg alone offers no provable privacy protection; practical deployment therefore requires threat-model-specific safeguards such as protected aggregation, access control, or formally calibrated differential privacy.
Owing to its intuitive visualization capability, state plane analysis (SPA) has been widely adopted to characterize the dynamic operating performance of LLC resonant converters, spurring the development of numerous SPA-based control strategies for such converters. However, the existing literature lacks a systematic review and quantitative analysis of these SPA-based control methods. This paper first elaborates on the core advantages of SPA that make it uniquely suitable for LLC converter control, including intuitive operational visualization, effective characterization of strong nonlinearities, analytical simplification via geometric principles, and multi-scale state characterization capability. On this basis, mainstream SPA-based control strategies—including optimal trajectory control (OTC), simplified OTC (SOTC), steady-state SPA-based feedforward control (SSPA-FFC), and averaged SPA-based OTC (ASPA-OTC)—are systematically categorized, with their fundamental working principles and implementation mechanisms elaborated in detail. This paper further analyzes the extended applications of SPA-based control in soft start-up, burst mode operation, synchronous rectification, and topology morphing, and proposes matched control combinations for each application scenario, establishing a practical implementation framework for SPA-based LLC converter control. Finally, this paper summarizes the research status of SPA-based control for LLC converters and outlines its key future development prospects.
The transition toward Software-Defined Vehicles has expanded the attack surface of the In-Vehicle Network, where the Controller Area Network (CAN) remains the de facto standard despite its lack of native security mechanisms. Cryptographic countermeasures such as AUTOSAR Secure Onboard Communication impose non-negligible overhead and cannot be retrofitted on legacy electronic control units, motivating lightweight Intrusion Detection System (IDS) as a practical alternative. This work proposes an orthogonal multi-layer IDS that analyzes complementary characteristics of each incoming CAN frame across the data-link and application layers. The first module, ML-Detector, employs a lightweight binary classifier that combines a logarithmic temporal feature, a one-hot encoding of nominal transmission periods, and the raw payload bytes. The second module, DBC-Verifier, deterministically verifies identifier whitelists, Data Length Code (DLC), and message periodicity constraints derived from the CAN database. The IDS firmware is partially generated via the MATLAB Embedded Coder toolchain, deployed on an automotive-grade NXP S32K3X4EVB-T172 board hosting an ARM Cortex-M7 core, and validated on VAGILETTA, a custom Hardware-in-the-Loop CAN testbed generating realistic traffic from eight ECUs over 56 CAN identifiers. Experimental results on five representative scenarios spanning DoS, black-/gray-box fuzzing, whitelist-aware payload spoofing, and a timing-compliant ECU-takeover condition show detection accuracies between 86.3% and 97.6%, while the complete IDS executes in only 38.84$\, \mathrm{\mu s}$ and occupies 3.12$\, \rm{kB}$ of Flash memory, demonstrating its suitability for real-time deployment on resource-constrained automotive ECUs.
