康明斯

Accurate identification of failure modes is essential for predictive maintenance of hydrogen-fueled solid oxide fuel cells (SOFCs). This study presents an electrochemical impedance-based failure mode diagnosis framework trained on experimentally generated degradation trajectories from twelve SOFCs subjected to four high-impact failure conditions: redox cycling, fuel starvation, short circuit, and thermal shock. Frequent electrochemical impedance spectroscopy measurements are converted into diagnostic signatures using distribution of relaxation times (DRT) analysis, and the resulting DRT-peak features and ohmic resistance are organized into labeled multivariate time-series windows for classification. The extracted DRT features are linked to underlying electrochemical phenomena within the cell, enabling failure modes to be interpreted through their impedance signatures rather than only through global power degradation. To benchmark diagnostic capability across different modeling paradigms, six classifiers are evaluated, including feature-based Random Forest (RF), XGBoost (XGB), and RF+XGB ensemble models, as well as sequence learners including Long Short-Term Memory, Temporal Convolutional Network, and an adapted attention-based Informer. Results demonstrate that temporal deep models provide improved balanced performance, with Informer achieving the strongest overall performance (Macro-F1 = 0.98). A sensitivity analysis further evaluates the repeatability of the impedance-derived features across repeated experiments for each failure mode. Finally, a computational cost analysis quantifies the performance-efficiency trade-off across model families.

The impact of mild hydrothermal aging (HTA) on low-temperature (150-200 degrees C) standard-SCR is investigated using transient response methods and transient kinetic analysis. We decouple the reduction and the oxidation half-cycles (RHC and OHC) of the standard-SCR redox mechanism to study them independently. While the RHC rates are essentially unaffected, OHC is inhibited by mild HTA. By equating the estimated rate expressions (both 2nd order in Cu sites, OHC with O2 as the sole oxidant), we predict exactly the steady-state low-temperature standard-SCR performance in terms of both NO conversion and bed-average Cu redox state, as well as the detrimental effect of mild HTA on the DeNOx efficiency. We also demonstrate that the DeNOx activity of the aged catalyst can be precisely restored by incrementing the oxygen partial pressure in proportion to the drop of the OHC rate constant, as predicted by our simple two-reaction model. These findings offer valuable insights into the design of next-generation urea-SCR exhaust gas aftertreatment (EGA) systems featuring enhanced cold-start performance and durability under aging conditions.

Improving diesel engine efficiency, reducing emissions, and enabling robust health monitoring have been critical research topics in engine modelling. While recent advancements in the use of neural networks for system monitoring have shown promising results, such methods often focus on component-level analysis, lack generalizability, and physical interpretability. In this study, we propose a novel hybrid framework that combines physics-informed neural networks (PINNs) with deep operator networks (DeepONet) to enable accurate and computationally efficient parameter identification in mean-value diesel engine models. Our method leverages physics-based system knowledge in combination with data-driven training of neural networks to enhance model applicability. Incorporating offline-trained DeepONets to predict actuator dynamics significantly lowers the online computation cost when compared to the existing PINN framework. To address the re-training burden typical of PINNs under varying input conditions, we propose two transfer learning (TL) strategies: (i) a multi-stage TL scheme offering better runtime efficiency than full online training of the PINN model and (ii) a few-shot TL scheme that freezes a shared multi-head network body and computes physics-based derivatives required for model training outside the training loop. The second strategy offers a computationally inexpensive and physics-based approach for predicting engine dynamics and parameter identification, improving computational efficiency over the existing PINN framework. Compared to existing health monitoring methods, our framework combines the interpretability of physics-based models with the flexibility of deep learning, offering substantial gains in generalization, accuracy, and deployment efficiency for diesel engine diagnostics.

Emerging power generation technologies such as solid oxide fuel cells (SOFCs) offer promising pathways for clean and efficient energy conversion. However, their material instability accelerates unexpected degradation, which remains a major barrier to large-scale commercialization. As a practical solution, control and diagnosis systems are integral to optimizing SOFCs' lifetime and efficiency in real-world operations. Degradation mechanisms induce nonlinear, time-varying patterns in running fuel cells, affecting the effectiveness of control and diagnosis strategies. This study introduces a data-driven machine learning framework for predicting SOFC performance under degradation. Four lab-scale SOFCs were subjected to accelerated degradation tests, generating diverse run-to-failure datasets. To capture the complex, nonstationary dynamics in these data, a dynamic neural network based on Kolmogorov-Arnold approximation theory (DKAN) is developed. DKAN employs univariate splines as learnable activation functions, hierarchically adapting low-dimensional functions to diverse nonlinearities and temporal patterns. Comparative experiments against state-of-the-art sequence models, including LSTM, TCN, WaveNet, DGRU, Informer, and ConvRec, show that DKAN achieves on average 30% lower prediction error (across RMSE, WAPE, and MASE) and 55% faster inference relative to the baselines, while demonstrating superior generalization to unseen degradation patterns. Furthermore, statistical analyses using the Friedman-Nemenyi and ANOVA-Tukey tests confirm the significance of DKAN's performance improvements across multiple datasets and metrics. These results highlight DKAN's potential as a lightweight and scalable solution for real-time SOFC diagnostics and control.

This study develops an aging-aware deep reinforcement learning framework for heavy-duty electrified powertrains, optimizing fuel consumption jointly with three physics-based degradation dynamics: battery calendar and cycling aging and generator thermal aging. The discrete–continuous actions are addressed using Physics-Aware Soft Actor–Critic (PA-SAC) agents, and training stability is enhanced by a Dynamic Programming (DP)-based warm start combining behavior cloning, expert replay-buffer seeding, and critic warmup. The framework is applied to a range-extender series-hybrid architecture, with agents trained on the Heavy Heavy-Duty Diesel Truck (HHDDT) cycle and evaluated on the training cycle and two unseen test cycles under a terminal state-of-charge (SOC) constraint of [0.5480, 0.5520]. Benchmarked across five seeds per case against DP optima, Twin Delayed Deep Deterministic Policy Gradient (TD3), and Proximal Policy Optimization with Lagrangian constraints (PPO-L), PA-SAC attains the tightest DP gap on the training cycle and lowest seed-to-seed variability. Aging objectives reshape the control strategy, reducing battery capacity degradation due to cycling aging by up to 91% relative to the fuel-only baseline and redistributing load between the battery and engine-generator subsystems under the aging-aware objective. The trained policy is deployed on an embedded controller target in a controller hardware-in-the-loop testbench, reproducing the training-time control behavior across twelve runs.