Index
Transfer learning Based Forecasting Of Heat Pump Energy Consumption Across Multiple Time Horizons
Die shape prediction using machine learning
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
Mutual Information-Based Segmentation for Unseen Domain Generalization in Digital Pathology
The introduction of automated slide scanners has facilitated the digitization of histopathological samples, enhancing the capabilities of traditional light microscopy by allowing the use of automated image analysis algorithms. Machine learning algorithms have demonstrated great potential in this regard by extrapolating learned characteristics from annotated datasets to unseen data, thus providing valuable assistance to pathologists in their diagnostic work. The performance of these models, however, can be significantly degraded by variations in image characteristics, including differences in scanners used for image acquisition, staining methods, resolution, illumination, and artifacts [1, 2]. These challenges highlight the difficulty of applying trained models across environments, necessitating domain adaption techniques.
Previous studies have already addressed color inconsistencies in histological samples, with calibration slides being one approach to resolving scanner-dependent variations [3]. Further notable pre-processing (-)/ training (⋆) techniques include:
– Data augmentation to simulate variability in the input data (e.g. domain-, spatial transformations) [4,5]
– Image-level domain adaption to align visual features across domains, mitigating distributional discrepancies, e.g. stain normalization to reduce inter-sample/ inter-scanner color variation [5,6]
– Multi-scale processing to capture features at different resolutions [2]
⋆ Heterogeneous dataset training to improve model generalization across multiple sources [7]
⋆ Transfer learning to utilize pre-trained models which is ideal for sparsely annotated data [2]
⋆ Domain-invariant feature learning to ensure robustness to scanner and staining variability [8,9], and in particular adversarial training to reinforce robustness against domain shifts [2]
⋆ Disentangled feature learning to isolate distinct underlying factors of data variations, compelling the network to learn shared statistical components across different domains [5]
This thesis investigates the applicability of a mutual information-based method for feature disentanglement [5] for cross-domain tumor segmentation in histopathology samples. By separating anatomical features from domain-specific variations, we aim for robust scanner-invariant segmentation performance. The objective is to enhance the generalizability of the network and enable direct application to unseen domains without adaptation.
The proposed work comprises the following work items:
– Literature review of device-induced variations in microscopy image data and state-of-the-art methods to address them
– Conceptualization and adaptation of mutual information-based segmentation [5] to address generalization for unseen domains in microscopy image data
– Exploration of targeted augmentation methods for addressing domain shifts in histopathology (e.g. stain augmentation [6])
– Exploration of suitable metrics for evaluating cross-domain generalization performance
– Documentation and presentation of the findings, documentation of code
[1] F. Wilm, M. Fragoso, C. A. Bertram, N. Stathonikos, M. Öttl, J. Qiu, R. Klopfleisch, A. Maier, K. Breininger, and M. Aubreville, “Multi-scanner canine cutaneous squamous cell carcinoma histopathology dataset,” in Bildverarbeitung für die Medizin 2023: Proceedings, German Workshop on Medical Image Computing, Braunschweig, July 2-4, 2023 (T. M. Deserno, H. Handels, A. Maier, K. Maier-Hein, C. Palm, and T. Tolxdorff, eds.), Informatik aktuell, Wiesbaden: Springer Fachmedien Wiesbaden, 2023.
[2] C. L. Srinidhi, O. Ciga, and A. L. Martel, “Deep neural network models for computational histopathology: A survey,” Medical Image Analysis, vol. 67, p. 101813, Jan. 2021.
[3] X. Ji, R. Salmon, N. Mulliqi, U. Khan, Y. Wang, A. Blilie, B. G. Pedersen, K. D. Sørensen, B. P. Ulhøi, R. Kjosavik, E. A. M. Janssen, M. Rantalainen, L. Egevad, P. Ruusuvuori, M. Eklund, and K. Kartasalo, “Physical Color Calibration of Digital Pathology Scanners for Robust Artificial Intelligence Assisted Cancer Diagnosis.”
[4] M. Balkenhol, N. Karssemeijer, G. J. S. Litjens, J. Van Der Laak, F. Ciompi, and D. Tellez, “H&E stain augmentation improves generalization of convolutional networks for histopathological mitosis detection,” in Medical Imaging 2018: Digital Pathology (M. N. Gurcan and J. E. Tomaszewski, eds.), (Houston, United States), p. 34, SPIE, Mar. 2018.
[5] Y. Bi, Z. Jiang, R. Clarenbach, R. Ghotbi, A. Karlas, and N. Navab, “MI-SegNet: Mutual Information-Based US Segmentation for Unseen Domain Generalization,” Feb. 2024. arXiv:2303.12649.
[6] M. Macenko, M. Niethammer, J. S. Marron, D. Borland, J. T. Woosley, Xiaojun Guan, C. Schmitt, and N. E. Thomas, “A method for normalizing histology slides for quantitative analysis,” in 2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, (Boston, MA, USA), pp. 1107–1110, IEEE, June 2009.
[7] M. Aubreville, F. Wilm, N. Stathonikos, K. Breininger, T. A. Donovan, S. Jabari, M. Veta, J. Ganz, J. Ammeling, P. J. Van Diest, R. Klopfleisch, and C. A. Bertram, “A comprehensive multi-domain dataset for mitotic figure detection,” Scientific Data, vol. 10, p. 484, July 2023.
[8] A. Moyes, “A Novel Method For Unsupervised Scanner-Invariance With DCAE Model.”
[9] M. W. Lafarge, J. P. W. Pluim, K. A. J. Eppenhof, P. Moeskops, and M. Veta, “Domain-adversarial neural networks to address the appearance variability of histopathology images,” 2017. arXiv:1707.06183.