Physical Knowledge inside Neural Networks for Anomaly Detection in Vehicle Signals

  • Subject:Physics-informed machine learning, anomaly detection in cyber-physical systems
  • Type:Masterarbeit
  • Date:ab 10 / 2026
  • Tutor:

    M. Sc. Luca Seidel

  • Zusatzfeld:

    luca.seidel@kit.edu

Physics-Based Knowledge in Neural Networks for Anomaly Detection in Vehicle Signals

Newton

Context

Vehicles are becoming software-defined systems: functions are increasingly delivered and updated independently of the hardware and redistributed across compute nodes at runtime. The coupling to the environment remains. Through sensors and actuators, software acts on physical processes. In open road traffic, its operating conditions cannot be fully captured at design time. Whether a sensor signal is plausible must therefore be assessed at runtime.

Two sources of knowledge with complementary weaknesses are available for this. Physical models are interpretable and verifiable, but limited to the operating conditions anticipated in advance. Data-driven anomaly detection also captures unforeseen patterns, but is hard to interpret. Physical knowledge is usually added around a network: as an extra input, a loss penalty or a downstream check. Approaches that embed prior knowledge into the network structure itself have so far relied on generic temporal invariants. Whether and how system-specific physical relations between signals can be embedded there remains open.

Initial reading

  • E. A. Lee, S. A. Seshia: Introduction to Embedded Systems: A Cyber-Physical Systems Approach, 2nd ed., MIT Press, 2017.
  • L. von Rueden et al., "Informed Machine Learning – A Taxonomy and Survey of Integrating Prior Knowledge into Learning Systems," in IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 1, pp. 614–633, Jan. 1, 2023, doi: 10.1109/TKDE.2021.3079836.
  • Z. Zamanzadeh Darban et al.: Deep Learning for Time Series Anomaly Detection: A Survey, ACM Computing Surveys 57(1), 2025.
  • S. Tuli, G. Casale, N. R. Jennings: TranAD: Deep Transformer Networks for Anomaly Detection in Multivariate Time Series Data, Proc. VLDB Endowment 15(6), 2022.
  • J. Xu, H. Wu, J. Wang, M. Long: Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy, ICLR, 2022.
  • S. Maleki, N. Pourmoazemi: Pi-Transformer: A Prior-Informed Dual-Attention Model for Multivariate Time-Series Anomaly Detection, Applied Soft Computing, 2026.

Targets

The goal is to design, implement and evaluate LSTM or Transformer architectures whose internal structure (e.g. attention, cell state, weight structure) embeds known physical relations. Choosing the backbones is part of the thesis.

  • Literature review on physics-informed and prior-informed architectures for time-series anomaly detection
  • Design of at least two mechanisms for embedding physical relations into the backbone, or implementation in two backbones
  • Evaluation on a synthetic dataset with known relations and injected faults, compared against (a) the same backbone without prior knowledge and (b) an approach with a generic prior (e.g. Pi-Transformer)
  • Systematic comparison of system-specific versus generic prior knowledge
  • Transfer to real vehicle bus signals and analysis per fault class

Requirements

  • Python and PyTorch
  • Solid knowledge of neural networks (LSTM/Transformer)
  • Independent working style