Data-Driven Approaches to Predict Material Failure and Analyze Material Models: Unterschied zwischen den Versionen

Aus SDQ-Institutsseminar
(Die Seite wurde neu angelegt: „{{Vortrag |vortragender=Martin Gauch |email=martin.gauch@student.kit.edu |vortragstyp=Bachelorarbeit |betreuer=Holger Trittenbach |termin=Institutsseminar/2017…“)
 
Keine Bearbeitungszusammenfassung
 
Zeile 5: Zeile 5:
|betreuer=Holger Trittenbach
|betreuer=Holger Trittenbach
|termin=Institutsseminar/2017-10-20 Zusatztermin
|termin=Institutsseminar/2017-10-20 Zusatztermin
|kurzfassung=Te prediction of material failure is useful in many industrial contexts such as predictive
|kurzfassung=Te prediction of material failure is useful in many industrial contexts such as predictive maintenance, where it helps reducing costs by preventing outages. However, failure prediction is a complex task. Typically, material scientists need to create a physical material model to run computer simulations. In real-world scenarios, the creation of such models is ofen not feasible, as the measurement of exact material parameters is too expensive. Material scientists can use material models to generate simulation data. Tese data sets are multivariate sensor value time series. In this thesis we develop data-driven models to predict upcoming failure of an observed material. We identify and implement recurrent neural network architectures, as recent research indicated that these are well suited for predictions on time series. We compare the prediction performance with traditional models that do not directly predict on time series but involve an additional step of feature calculation. Finally, we analyze the predictions to fnd abstractions in the underlying material model that lead to unrealistic simulation data and thus impede accurate failure prediction. Knowing such abstractions empowers material scientists to refne the simulation models. The updated models would then contain more relevant information and make failure prediction more precise.
maintenance, where it helps reducing costs by preventing outages. However, failure prediction is a complex task. Typically, material scientists need to create a physical material
model to run computer simulations. In real-world scenarios, the creation of such models
is ofen not feasible, as the measurement of exact material parameters is too expensive.
Material scientists can use material models to generate simulation data. Tese data sets
are multivariate sensor value time series. In this thesis we develop data-driven models
to predict upcoming failure of an observed material. We identify and implement recurrent neural network architectures, as recent research indicated that these are well suited
for predictions on time series. We compare the prediction performance with traditional
models that do not directly predict on time series but involve an additional step of feature
calculation.
Finally, we analyze the predictions to fnd abstractions in the underlying material model
that lead to unrealistic simulation data and thus impede accurate failure prediction. Knowing such abstractions empowers material scientists to refne the simulation models. Te
updated models would then contain more relevant information and make failure prediction more precise.
}}
}}

Aktuelle Version vom 6. Oktober 2017, 10:08 Uhr

Vortragende(r) Martin Gauch
Vortragstyp Bachelorarbeit
Betreuer(in) Holger Trittenbach
Termin Fr 20. Oktober 2017
Vortragsmodus
Kurzfassung Te prediction of material failure is useful in many industrial contexts such as predictive maintenance, where it helps reducing costs by preventing outages. However, failure prediction is a complex task. Typically, material scientists need to create a physical material model to run computer simulations. In real-world scenarios, the creation of such models is ofen not feasible, as the measurement of exact material parameters is too expensive. Material scientists can use material models to generate simulation data. Tese data sets are multivariate sensor value time series. In this thesis we develop data-driven models to predict upcoming failure of an observed material. We identify and implement recurrent neural network architectures, as recent research indicated that these are well suited for predictions on time series. We compare the prediction performance with traditional models that do not directly predict on time series but involve an additional step of feature calculation. Finally, we analyze the predictions to fnd abstractions in the underlying material model that lead to unrealistic simulation data and thus impede accurate failure prediction. Knowing such abstractions empowers material scientists to refne the simulation models. The updated models would then contain more relevant information and make failure prediction more precise.