Deep Learning and System Identification

Ljung, Lennart; Andersson, Carl; Tiels, Koen; Schön, Thomas B.

doi:10.1016/j.ifacol.2020.12.1329

Cited by 122 publications

(61 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Therefore, we choose n d 5 as a fixed value for two reasons: first, it guaranteed a robust fitting of the testing data even when the size of the training set was reduced, thus avoiding overfitting issues (Table 3). On the other hand, n d 5 allowed to obtain better interpretable models with respect to system identification approaches, e.g., neural networks or deep learning, for which the goodness of fitting is favored with respect to the model interpretability (Ljung et al, 2020).…”

Section: Discussionmentioning

confidence: 99%

Neuromuscular Control Modelling of Human Perturbed Posture Through Piecewise Affine Autoregressive With Exogenous Input Models

Tigrini

Verdini

Maiolatesi

et al. 2022

Front. Bioeng. Biotechnol.

View full text Add to dashboard Cite

In this study, the neuromuscular control modeling of the perturbed human upright stance is assessed through piecewise affine autoregressive with exogenous input (PWARX) models. Ten healthy subjects underwent an experimental protocol where visual deprivation and cognitive load are applied to evaluate whether PWARX can be used for modeling the role of the central nervous system (CNS) in balance maintenance in different conditions. Balance maintenance is modeled as a single-link inverted pendulum; and kinematic, dynamic, and electromyography (EMG) data are used to fit the PWARX models of the CNS activity. Models are trained on 70% and tested on the 30% of unseen data belonging to the remaining dataset. The models are able to capture which factors the CNS is subjected to, showing a fitting accuracy higher than 90% for each experimental condition. The models present a switch between two different control dynamics, coherent with the physiological response to a sudden balance perturbation and mirrored by the data-driven lag selection for data time series. The outcomes of this study indicate that hybrid postural control policies, yet investigated for unperturbed stance, could be an appropriate motor control paradigm when balance maintenance undergoes external disruption.

show abstract

Section: Discussionmentioning

confidence: 99%

Neuromuscular Control Modelling of Human Perturbed Posture Through Piecewise Affine Autoregressive With Exogenous Input Models

Tigrini

Verdini

Maiolatesi

et al. 2022

Front. Bioeng. Biotechnol.

View full text Add to dashboard Cite

show abstract

“…18). The principal idea is to seek a system identification operator [29] that corrects the amplitude residuals in bulk quantities which the CRAN architecture develops. We achieve this by learning the initial 300 performance steps and forecasting all 500 steps.…”

Section: Near-wake Dynamics Using the Cnn-rnn Drivermentioning

confidence: 99%

A hybrid partitioned deep learning methodology for moving interface and fluid-structure interaction

Gupta¹,

Jaiman²

2021

Preprint

View full text Add to dashboard Cite

In this work, we present a hybrid physics-based deep learning (DL) framework for handling moving interfaces and predicting fluid-structure interaction (FSI). Using the discretized Navier-Stokes (NS) in the Arbitrary Lagrangian-Eulerian (ALE) reference frame, we generate full-order flow snapshots and point-cloud displacements as target physical data for the learning and inference of coupled fluid-structure dynamics. This integrated operation of the physics-based modeling with the DL-based reduced-order model (DL-ROM) makes our framework hybrid. The proposed multi-level framework is composed of two data-driven physics-DL drivers that predict unsteady flow and track the moving point cloud displacements respectively, while synchronously exchange the force information at the physical interface. The first component of our proposed framework relies on the proper orthogonal decomposition-based recurrent neural network (POD-RNN) as a semi-supervised procedure to infer the point cloud ALE description. This model essentially relies on the POD basis modes to reduce dimensionality and evolving them in the time domain. The second component utilizes the convolution-based recurrent autoencoder network (CRAN) as a self-supervised DL procedure to predict the nonlinear flow dynamics at static Eulerian probes. We introduce these probes as spatially structured query nodes in the moving point cloud to resolve the field Lagrangian to Eulerian conflict as well as conveniently train the CRAN driver. We design a novel snapshot-field transfer and load recovery (FTLR) algorithm to optimally select these Eulerian probes. They are selected in such a way that the two components (i.e., POD-RNN and CRAN) are constrained at the interface to recover bulk force quantities. These hybrid physics-DL drivers ultimately rely on recurrent neural networks to evolve the low-dimensional state. A popular prototypical fluid-structure interaction problem of flow past a freely oscillating cylinder is selected to test the efficacy of the proposed hybrid DL-ROM methodology. The framework tracks the interface description accurately and predicts highly non-linear wake dynamics for nearly 500 time-steps with limited input demonstrator within excellent to a reasonable accuracy. These results motivate us to further explore the application of this hybrid framework for the concept of digital twinning of engineering systems, especially those involving moving boundaries and fluid-structure interactions.

show abstract

“…Although neural networks that are based on convolutional or recurrent layers have inherent capabilities to model dynamic systems, only a limited amount of work omitted the autoregression. It has been shown, that RNNs behave like an NLSS (Ljung et al, 2020) and an RNN has been applied non-autoregressively to a synthetic dataset (Gonzalez and Yu, 2018). The authors of the present paper applied nonautoregressive TCNs and gated recurrent units (GRUs) to an inertial measurement-based sensor fusion task, outperforming state-of-the-art domain-specific sensor fusion methods (Weber et al, 2020).…”

Section: Introductionmentioning

confidence: 98%

“…Most current neural network-based system identification methods still are NARX variants with different neural network architectures as nonlinearities. In related work, a multitude of system identification methods that are based on autoregressive neural networks has been proposed using multilayer perceptrons (MLPs) (Shi et al, 2019), cascaded MLPs (Ljung et al, 2020), convolutional neural networks (Lopez and Yu, 2017), TCNs (Andersson et al, 2019), and recurrent neural networks (Kumar et al, 2019) with promising results.…”

Section: Introductionmentioning

confidence: 99%

Non-Autoregressive vs Autoregressive Neural Networks for System Identification

Weber¹,

Gühmann²

2021

Preprint

View full text Add to dashboard Cite

The application of neural networks to non-linear dynamic system identification tasks has a long history, which consists mostly of autoregressive approaches. Autoregression, the usage of the model outputs of previous time steps, is a method of transferring a system state between time steps, which is not necessary for modeling dynamic systems with modern neural network structures, such as gated recurrent units (GRUs) and Temporal Convolutional Networks (TCNs). We compare the accuracy and execution performance of autoregressive and non-autoregressive implementations of a GRU and TCN on the simulation task of three publicly available system identification benchmarks. Our results show, that the non-autoregressive neural networks are significantly faster and at least as accurate as their autoregressive counterparts. Comparisons with other state-of-the-art black-box system identification methods show, that our implementation of the non-autoregressive GRU is the best performing neural network-based system identification method, and in the benchmarks without extrapolation, the best performing black-box method.

show abstract

Deep Learning and System Identification

Cited by 122 publications

References 9 publications

Neuromuscular Control Modelling of Human Perturbed Posture Through Piecewise Affine Autoregressive With Exogenous Input Models

Neuromuscular Control Modelling of Human Perturbed Posture Through Piecewise Affine Autoregressive With Exogenous Input Models

A hybrid partitioned deep learning methodology for moving interface and fluid-structure interaction

Non-Autoregressive vs Autoregressive Neural Networks for System Identification

Contact Info

Product

Resources

About