1 of 23

Effective State-Space Models Estimation for 2D Flows Using Convolutional Neural Networks in a SUBNET Approach

Gerben I. Beintemaa, Roland Totha,b , Maarten Schoukensa

aDepartment of Electrical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands

bSystems and Control Laboratory, Institute for Computer Science and Control, Budapest, Hungary.

emails: g.i.beintema@tue.nl, m.schoukens@tue.nl, r.toth@tue.nl

1

2 of 23

Spatial Systems

2

PDE simulation describe many systems accurately

Cameras are cheap

3 of 23

Analysis and Control of spatial systems

3

High dimensional/spatial system are challenging to analysis and control

Analysis/Design

Difficult due to high dimensional data

MPC

PDE simulations are often much slower than real time

Control synthesis

Applied up to

about 20 states

Strategy: Dimension Reduction

4 of 23

Model order reduction two approaches

4

System

Collect high Dim. data

High Dim.

IO data

Dim.

Reduction

Low Dim.

IO data

System Identification

Low Dim. model

Loss of information potentially important to system identification

PDE

System

Dim.

Reduction

Low Dim.

System

  1. Difficult for nonlinear system
  2. Impossible for cameras

Approach 1:

Approach 2:

5 of 23

This presentation

5

High Dim.

System

Collect data

High Dim.

data

Dim.

Reduction

Low Dim.

data

System Identification

Low Dim. model

This presentation

SUBNET method for low dim. data

SUBNET

method for

high dim. data

Add some AI

6 of 23

Prediction/Simulation Error Approach

6

Short comings:

The Dataset:

The Model:

The Loss function:

 

Gradient and value Explosions

Local Minima

Non-smooth cost

Many nice theory properties and can provide great models!

7 of 23

SUBNET approach[1]

7

 

The Datasets:

The Model:

The Loss function:

 

Explosions Reduced

Local Minima Reduced

Smoothed cost

Advantages:

[1] Beintema, Gerben I., Maarten Schoukens, and Roland Tóth. "Deep subspace encoders for nonlinear system identification." Automatica 156 (2023): 111210.

8 of 23

Construction Short Datasets

8

Extract short datasets

from full dataset

Allow for overlap:

More sample efficient[1]

[1] Beintema, Gerben I., Maarten Schoukens, and Roland Tóth. "Deep subspace encoders for nonlinear system identification." Automatica 156 (2023): 111210.

9 of 23

SUBNET Initial states

9

 

Option 2:

State Encoder

  1. Fixed complexity
  2. Theoretically: an approximate SUBspace reconstructability map

 

 

10 of 23

SUBNET Approach for NL state-space

10

Excellent (benchmark) results using MLPs

11 of 23

Extend SUBNET to high dim. 2D Data

11

 

high 2D Dim.

data

SUBNET identification

Low dim. State-space model

12 of 23

Main idea: Progressive upscale!

12

Their model structure:

  • Latent -> Small image (MLP)
  • Progressively upscale image (multiple CNNs)
  • Feature size decreases as image size increases

Take aways:

  • Many small steps > �few large steps
  • Avoid information bottlenecks
  • Avoid high function complexity

Karras, Tero, et al. "Progressive growing of GANS for improved quality, stability, and variation." arXiv preprint arXiv:1710.10196 (2017).

13 of 23

 

13

Stride = 2

Stride = 1

Downscaling Network

14 of 23

 

14

Upscaling Network

15 of 23

Computational Properties

15

Downscaling Network

Upscaling Network

Each upscale CNN:

Image W & H * 2

Feature size / 2

Each downscale CNN:

Image W & H / 2

Feature size * 1.5

Linear Scaling

Max image size:

About 250 by 250

16 of 23

Benchmark: Wake Flow Oscillating Cylinder

16

 

Decuyper, J., De Troyer, T., Tiels, K., Schoukens, J., & Runacres, M. C. (2020). A nonlinear model of vortex-induced forces on an oscillating cylinder in a fluid flow. Journal of Fluids and Structures, 96, 103029.

17 of 23

SUBNET and model structure

17

SUBNET parameters:

None of these are specify sensitive parameters

Others settings:

CNN Kernel size: 3

Activation: tanh

Normalization: Input and Output

MLPs: 2 hidden layers 64 nodes

Optimizer: Adam optimizer

Compute: 1 hour compute on GPU (1660)

18 of 23

Important: Good Early Stopping

18

Train SUBNET loss

Val. Sim. Error

Roll back to best validation error

Epochs

19 of 23

Results: Wake Flow Oscillating Cylinder

19

1.6M parameters

200K parameters

20 of 23

The Numbers

20

Model Inference: < 1 sec

Original PDE: > 1 day

Training: Few hours on a GPU

Errors:

About 8% NRMS

(excluding autonomous part)

21 of 23

Result: Controlled Thermal Convection

21

Gerben Beintema, Data–driven Learning of Nonlinear Dynamic Systems: A Deep Neural State–Space Approach, PhD Thesis 2024

22 of 23

Results: Camera on Unbalanced disk

22

23 of 23

Take aways

23

Our Aim:

We extended the SUBNET to high dim. 2D Data

Parameterization: Many small steps >

Few big steps

Progressive CNN up/downscaling

scales well!

Results:

Modelled both PDE data and camaras with great success