Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Learning of Dynamic Systems using System Identification Toolbox(TM)

Tianyu Dai, Khaled Aljanaideh, Rong Chen, Rajiv Singh, Alec Stothert, Lennart Ljung

TL;DR

The toolbox contains several other enhancements that deepen its integration with the state-of-art machine learning techniques, leverage auto-differentiation features for state estimation, and enable a direct use of raw numeric matrices and timetables for training models.

Abstract

MATLAB(R) releases over the last 3 years have witnessed a continuing growth in the dynamic modeling capabilities offered by the System Identification Toolbox(TM). The emphasis has been on integrating deep learning architectures and training techniques that facilitate the use of deep neural networks as building blocks of nonlinear models. The toolbox offers neural state-space models which can be extended with auto-encoding features that are particularly suited for reduced-order modeling of large systems. The toolbox contains several other enhancements that deepen its integration with the state-of-art machine learning techniques, leverage auto-differentiation features for state estimation, and enable a direct use of raw numeric matrices and timetables for training models.

Deep Learning of Dynamic Systems using System Identification Toolbox(TM)

TL;DR

The toolbox contains several other enhancements that deepen its integration with the state-of-art machine learning techniques, leverage auto-differentiation features for state estimation, and enable a direct use of raw numeric matrices and timetables for training models.

Abstract

MATLAB(R) releases over the last 3 years have witnessed a continuing growth in the dynamic modeling capabilities offered by the System Identification Toolbox(TM). The emphasis has been on integrating deep learning architectures and training techniques that facilitate the use of deep neural networks as building blocks of nonlinear models. The toolbox offers neural state-space models which can be extended with auto-encoding features that are particularly suited for reduced-order modeling of large systems. The toolbox contains several other enhancements that deepen its integration with the state-of-art machine learning techniques, leverage auto-differentiation features for state estimation, and enable a direct use of raw numeric matrices and timetables for training models.
Paper Structure (12 sections, 1 equation, 11 figures)

This paper contains 12 sections, 1 equation, 11 figures.

Table of Contents

  1. Introduction
  2. Neural State-Space Models
  3. Example: Black-Box Model of SI Engine Torque Dynamics
  4. Example: Feature Reduction Using Auto-Encoders
  5. Neural Networks for Nonlinear ARX and Hammerstein-Wiener models
  6. Example: Black-box Modeling of Robot Arm Dynamics
  7. Feature Selection and Dictionary-Based Learning
  8. Example: Feature Selection for IC Engine Model
  9. Other Improvements
  10. Use of Auto-Differentiation in Extended Kalman Filters
  11. Time-Domain Data Format
  12. Final Comments

Figures (11)

  • Figure 1: Possible configurations of the state transition function of a Neural State Space model. Top: Basic (default) configuration. Bottom: Configuration using auto-encoder to reduce or increase the latent state dimension.
  • Figure 2: SI Engine model.
  • Figure 3: SI Engine data.
  • Figure 4: SI Engine system: Fit score on validation data using an neural state-space model.
  • Figure 5: Two tank system.
  • ...and 6 more figures