best practices and use cases for data-driven engineering...2020/12/02  · data understanding:...

22
Best practices and use cases for data-driven engineering Dr. Stefan Suwelack [email protected]

Upload: others

Post on 27-Jan-2021

0 views

Category:

Documents


0 download

TRANSCRIPT

  • Best practices and use cases for

    data-driven engineering

    Dr. Stefan Suwelack [email protected]

  • Present use

    cases by role ?Test

    SimulationDesign

    PLM

    2© Renumics GmbH

  • Use horizontal

    categories Surrogate models

    Data exploration

    & understanding

    Generative

    design

    Process

    automation

    3© Renumics GmbH

  • 5

    Data understanding: Assessment of DoEs

    Example: Crash simulation

    Improve robustness

    Understand optimality criteria

    Understand critical correlations

    © Renumics GmbH

    © Constantin Diez, Lasso Engineering

    Classic method:

    Principal component analysis

    (PCA)

  • Deep autoencoders

    & representation

    learning

    Advantages:

    1. Geometry independence

    2. Generalized task-specific

    similarity measuresembedding

    6© Renumics GmbH

  • Data understanding: Assessment of test data

    Example: Acoustic measurement data

    Find outliers and reference points

    Identify root causes

    Find errors in test setup

    Classic method:

    Spectral features + shallow ML

    DL methods:

    CNN autoencoder, LSTMs

    7© Renumics GmbH

  • Process automation: Geometry processing

    Recognition and measurement of

    complex geometric features

    Time saving: >70%

    Assessment of collisions (packaging /

    DMU)

    Time savings: > 50%

    8© Renumics GmbH

  • Process automation: Simulation assessment

    Assessment of simulation results for

    NVH

    Mode classification >95% accuracy

    Simulation monitoring

    Reduction of simulation time, significant error

    reduction

    9© Renumics GmbH

  • Surrogate modeling

    Example: Digital twin of turbine blade

    Speed-up

    Efficient optimization loops

    Digital Twin / IOT applications

    © Dynardo GmbH

    Classic methods (examples):

    Response surface models (parametric),

    Proper orthogonal decomposition

    (fields)

    10© Renumics GmbH

  • Deep learning

    based surrogate

    modeling

    Advantages:

    1. Generalizes over different

    simulation settings

    2. Geometry can be

    parameter – free

    Big trend: Physics-informed

    neural networks© NVIDIA 2020

    11© Renumics GmbH

  • Generative

    Modeling

    Classic methods:

    1. Parametric models

    2. Topology optimization +

    reverse engineering +

    manufacturing rules

    12© Renumics GmbH

  • Generative

    Modeling with DL

    Advantages:

    Automatic selection of implicit

    parameters

    Big challenge: Dataset

    selection and extrapolation

    © Autodesk 2017

    13© Renumics GmbH

  • Best practices for

    successful ML

    projects

    1. Define business case

    2. Understand available data

    3. Build data strategy and tooling

    4. Define user interaction with ML-tool

    5. Prototype algorithms

    6. Prototype process integration

    7. Build trust with users

    8. Roll-out and operate

    14© Renumics GmbH

  • Building and

    curating datasets

    Historic Data

    + already available - messy

    + distribution well

    captured

    - siloed

    Synthetic Data

    + standardized setups

    and formats

    -expensive to obtain

    + lower bureaucratic

    obstacles

    -Difficult

    parameterization

    - biased datasets

    16© Renumics GmbH

  • Dataset curation: Similarity map examples I

    Simulationsrun by intern

    High modelerror: Noconvergence

    17© Renumics GmbH

  • Dataset curation: Similarity map examples III

    Difficultexamples

    Bad annotations?

    19© Renumics GmbH

  • Similarity

    maps

    Collaborative ML-model development:

    Discuss relevant data properties with

    domain experts

    Annotate quickly and reliably

    Discuss model performance

    Build user trust

    Analysis of complex engineering data:

    DoE workflows: Simulation result

    assessment

    Test data: Find points of interest

    20© Renumics GmbH

  • Powerful filters

    Interactive similarity map

    Highly configurabledetailed view

  • Renumics

    Spotlight

    Fast data exploration & ML-model design

    Curate datasets for ML training

    Understand large amounts of simulation

    data quickly

    Analyze test data and optimize test setups

    Modular integration into existing tools

    Easily integrates into Python-based

    workflows

    Can be enhanced with Renumics

    Backstage: Customized notebook for data

    science beginners

    Uses open standards to integrate into an

    open ecosystem of ML-tools

    22© Renumics GmbH

  • We help great engineers to

    understand their data

    Dr. Stefan Suwelack [email protected]

    Daniel Klitzke [email protected]

    Steffen Slavetinsky [email protected]