Index

Deep Learning for Cone-Beam CT Field-of-View Extension

Deep Learning Approaches for Dynamic Modeling of Hydrogen Fuel Cells in Hybrid Railway Vehicles

Motivation

Hydrogen is increasingly recognized as a promising clean energy carrier, particularly in the transportation sector, due to its potential for long driving ranges, rapid refueling, and reduced environmental footprint. In railway applications, hydrogen fuel cell hybrid trains are being developed for non-electrified routes. These vehicles integrate several interdependent subsystems for energy generation (fuel cell), storage (traction battery), and consumption (traction motors and auxiliary loads). Their performance, cost efficiency, and lifetime strongly depend on system-level design and operational strategies, which must be evaluated and optimized through accurate simulations.

A central challenge lies in the modeling and control of the fuel cell system, which is a nonlinear, dynamic, and degradation-prone component. Accurate models are essential not only for system design and optimization but also for the development of advanced control strategies that enhance efficiency, minimize hydrogen consumption, and prolong component lifetime.

State of the Art

Conventional fuel cell simulation models are typically based on simplified physics-based equations combined with laboratory-derived efficiency and performance maps [8] [7]. While suitable for steady-state operation, these maps fail to capture dynamic transitions, transient behavior, and degradation effects observed in real-world railway applications. Consequently, their predictive accuracy in realistic duty cycles remains limited.

Recent advances in data-driven modeling, including machine learning provide promising alternatives for capturing nonlinear system dynamics and adapting to varying conditions [4][5][6]. However, these approaches have not yet been systematically applied and validated for railway fuel cell hybrid systems using real-world operational data.

Furthermore, in industrial context, the models are trained once and applied for multiple vehicles of the same type. This neglects the fact that each vehicle has slightly different physical characteristics, e.g. due to production variance and different degradation state.
Recently, transfer methods have been published which enable the pretraining of a model based on data from one device and fine-tuning the model to another similar device [3]. Furthermore, methods for layer freezing enable the fine-tuning of selected neural network parts [1][2].

Planned Tasks

This thesis aims to develop data-driven models for hydrogen fuel cells in hybrid railway vehicles. The work is structured into two main parts:

• Processing and analysis of measurement data collected from real vehicles under dynamic operating conditions.
• Development of machine learning-based fuel cell models using neural network architectures (e.g., CNN, LSTM, or hybrid approaches).

• Exploration of a two-stage training approach. The first stage involves a “base training” using high-resolution, accurate sensor data from a test vehicle. The second stage involves an “adaptation training” using coarser data, which are available on all vehicles. During adaptation, parts of the neural network that capture fast dynamics may be frozen, while components modeling slower dynamics are fine-tuned to match specific vehicle characteristics.
• Quantitative validation and statistical benchmarking against conventional physics-based and map-based fuel cell models.

References

[1]   S. Li, G. Yuan, Y. Dai, Y. Zhang, Y. Wang, and X. Tang, “SmartFRZ: An efficient training framework using attention-based layer freezing,” arXiv, 2024. [Online]. Available: https://arxiv.org/abs/2401.16720

[2]   C. G. Krishnanunni and T. Bui-Thanh, “An adaptive and stability-promoting layerwise training approach for sparse deep neural network architecture,” arXiv, 2024. [Online]. Available: https://arxiv.org/abs/2211.06860

[3]   L. A. Briceno-Mena, J. A. Romagnoli, and C. G. Arges, “PemNet: A transfer learning-based modeling approach of high-temperature polymer electrolyte membrane electrochemical systems,” Industrial & Engineering Chemistry Research, vol. 61, no. 9, pp. 3350–3357, Feb. 2022, doi: 10.1021/acs.iecr.1c04237.

[4] Xuezhao Zhang, Zijie Chen, Wenxiao Wang, and Xiaofen Fang. Prediction method of phev driving energy consumption based on the optimized cnn bilstm attention network. Energies, 17(12), 2024.

[5] Pedro Lara-Benitez, Manuel Carranza-Garcia, and Jose C. Riquelme. An experimental review on deep learning architectures for time series forecasting. International Journal of Neural Systems, 31(03):2130001, February 2021.

[6] Alex Sherstinsky. Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm) network. Physica D: Nonlinear Phenomena, 404:132306, 2020.

[7] Petrone R., et al. GREY-BOX MODELLING FOR PEM FUEL CELL MONITORING

[8] Kapetanović, M., Núñez, A., van Oort, N., & Goverde, R. M. P. (2023). Energy model of a fuel cell hybrid-electric regional train in passenger transport service and vehicle-to-grid applications. Journal of Rail Transport Planning & Management, 28, Article 100415. DOI: 10.1016/j.jrtpm.2023.100415.

Prediction of CBCT Image Quality based on Fluoro Scout Images

Investigating DETR Architectures as Implant Detection Model for Metal Artifact Avoidance Application

Investigating YOLO Architectures as Implant Detection Model for Metal Artifact Avoidance Application

Evaluation and Fusion of Vision-Language and Computer Vision Models for On-road Scenario Extraction in Autonomous Vehicles

Agentic Radiology Report Generation

Large Language Models for Modified Frenchay Dysarthria Assessment Reports from Parkinson’s Speech: Model Choice and Prompting Effects

Advanced nnU-Net Ensemble Techniques for Lung Nodule Segmentation

This thesis outlines a comprehensive research program aimed at advancing medical
image segmentation through enhanced nnU-Net ensemble methodologies. Building upon
substantial experimental results that demonstrate significant improvement over existing
approaches, the proposed research addresses critical gaps in current medical imaging AI
capabilities.  The research aims to establish new performance standards for automated lung
nodule detection with current achievements of 0.84 dice coefficient representing a 29.2%
improvement over SAM baseline and approaching clinical utility thresholds. Future work will
focus on Vision Transformer integration with nnU-Net architectures, generalization validation
across additional lung imaging datasets, and clinical deployment optimization. The expected
outcomes include significant academic contributions through peer-reviewed publications,
practical clinical applications with potential for real-world healthcare impact, and establishment
of open-source implementations for research community adoption.

Development of a Local LLM Agent System for Clinical Expert Support and Automation in MRI Planning for Radiation Therapy

If you are interested, please contact fabian.wagner@fau.de