M1 Sense

Systematic Over-Instrumentation

M1 decides which sensors and actuators to add while a process is immature, and which to remove once it matures.

Illustration. The green field shows where measurements would reveal the most about wrinkle formation. Candidate sensors on the sheet and grippers are ranked and selected, and two of them are removed as the process matures.

Work packages

Results of the first funding phase, by work package.

  1. M1.1 Optimal sensor and actuator placement and parameterisation

    Using a high-fidelity forming simulation from T2, a Bayesian optimisation developed with M3 finds gripper configurations that improve product quality and decides where to place an additional gripper. A second method places additional or mobile sensors for fault diagnosis under model mismatch. Validation on the real system is planned.

    Three simulated formed parts seen from above, with curvature in colour and the gripper positions marked around each part. Enlarge figure
    Simulated forming results, with curvature in colour. (a) Eight grippers in a configuration found by Bayesian optimisation. (b) Improved result with an additional, locally optimised gripper. (c) The additional gripper in a less effective position. Figure: After Döhner et al. 2026.

    Publications

  2. M1.2 Discovery of latent variables

    Causal root-cause analysis combines anomaly detection with expert-defined causal graphs and process-step information, and a counterfactual extension explains why one root cause is more likely than another. Other methods learn causal structure from wavelet-based soft interventions, evaluate learned causal graphs with a normalised causal edit distance, and learn dynamic structural causal mechanisms from time series.

    Architecture diagram of the dynamic structural causal mechanisms method, from observational time series through a GRU encoder and graph-constrained neural mechanisms to counterfactual inference and optimisation. Enlarge figure
    Overview of the dynamic structural causal mechanisms (DSCM) method. A GRU and environment embeddings encode the observational time series, and the encoding is combined with graph-constrained neural mechanisms to learn dynamic causal relationships. Figure: KI-FOR 5339.
  3. M1.3 Sensitivity and influence analysis

    A method combining causal discovery, causal inference and temporal neural networks selects the variables that causally influence the outcome and quantifies their effect with the conditional average treatment effect. These variables form the basis for a minimal set of relevant sensors.

  4. M1.4 Application and validation

    Patterns printed on the laminate before forming make defects such as folds visible afterwards, and methods to detect such folds automatically are under development. A dedicated test rig measures how low-frequency tool vibrations influence forming. Transferring the Bayesian optimisation of the gripper configuration to the real forming process is planned.

About the subproject

Early on, generous instrumentation is the fastest way to understand a new process. Later, many of the sensors are no longer needed, and the process should run with fewer, cheaper ones.

Principal investigators
Researchers
  • Hannah Decker
  • Frank Döhner
  • Zahra Nasrollah
  • Shahenda Youssef
  • Georg Zeeb
Former members
  • Dr. Josephine Rehak
Host institute
Institute of Vehicle System Technology (FAST), KIT
People in M1

Publications 10