Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Experiment Video

Updated: Jun 30, 2026

Design and Application of a Fault Detection Method Based on Adaptive Filters and Rotational Speed Estimation for an Electro-Hydrostatic Actuator
06:45

Design and Application of a Fault Detection Method Based on Adaptive Filters and Rotational Speed Estimation for an Electro-Hydrostatic Actuator

Published on: October 28, 2022

An algorithmic approach to adaptive state filtering using recurrent neural networks.

A G Parlos1, S K Menon, A Atiya

  • 1Department of Mechanical Engineering, Texas A&M University College Station, TX 77843, USA. a-parlos@tamu.edu

IEEE Transactions on Neural Networks
|February 6, 2008
PubMed
Summary
This summary is machine-generated.

Related Concept Videos

State Space Representation01:27

State Space Representation

The frequency-domain technique, commonly used in analyzing and designing feedback control systems, is effective for linear, time-invariant systems. However, it falls short when dealing with nonlinear, time-varying, and multiple-input multiple-output systems. The time-domain or state-space approach addresses these limitations by utilizing state variables to construct simultaneous, first-order differential equations, known as state equations, for an nth-order system.
Consider an RLC circuit, a...

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Rationale, study design and methodology of the LANDMARC trial: a 3-year, pan-India, prospective, longitudinal study to assess management and real-world outcomes of diabetes mellitus.

Diabetic medicine : a journal of the British Diabetic Association·2019
Same author

Prevalence and characteristics of cigarillo smoking in Canada: results from the Canadian Tobacco Use Monitoring Survey.

Public health·2018
Same author

Autoimmune hypophysitis: a single centre experience.

Singapore medical journal·2009
Same author

Congenital adrenal hyperplasia 11beta-hydroxylase deficiency: two cases managed with bilateral adrenalectomy.

Singapore medical journal·2009
Same author

Sequential liquid-liquid extraction graphite furnace atomic absorption spectrometric (GFAAS) and spectrophotometric determination of zirconium(IV) with calixarene hydroxamic acid.

Talanta·2008
Same author

New results on recurrent network training: unifying the algorithms and accelerating convergence.

IEEE transactions on neural networks·2008

This article introduces new computational methods for tracking the internal states of complex, changing systems when the underlying mathematical rules are not fully understood. By using neural networks to learn these rules, the researchers created flexible filters that perform well even when noise levels or system behaviors are unpredictable. These tools offer a robust alternative to traditional mathematical approaches, providing higher accuracy and better stability in challenging environments.

Area of Science:

  • Computational intelligence and recurrent neural networks in control theory
  • Nonlinear dynamics and adaptive state filtering systems analysis

Background:

Engineers often struggle to track hidden variables within complex systems when the governing mathematical laws remain undefined. Prior research has shown that traditional estimation techniques rely heavily on precise models of system behavior. That uncertainty drove the development of methods that can learn dynamics directly from observed data. No prior work had resolved how to maintain stability when noise statistics are completely unknown. This gap motivated the exploration of machine learning architectures for state estimation tasks. It was already known that standard linear approaches fail to capture the nuances of non-linear environments. Researchers have long sought ways to bypass the need for explicit system equations in real-time tracking. This article addresses these challenges by applying neural network approximations to dynamic state filtering problems.

Purpose Of The Study:

The aim of this research is to present practical algorithms for adaptive state filtering in nonlinear dynamic systems where the state equations are unknown. The study addresses the challenge of tracking hidden variables without relying on precise mathematical models. Researchers seek to demonstrate that neural networks can constructively approximate these missing dynamics. The motivation stems from the limitations of traditional estimation techniques when faced with unpredictable noise statistics. This work explores how machine learning can provide a more flexible framework for real-time state tracking. The authors intend to compare their proposed neural-based filters against established extended Kalman filter methods. By evaluating both off-line and online learning stages, the team investigates the potential for improved estimation accuracy. This effort aims to establish a robust methodology for state estimation in environments where system behavior is difficult to define analytically.

Keywords:
state estimationmachine learning controlKalman filter alternativesadaptive filtering

Frequently Asked Questions

The researchers propose a two-step prediction-update mechanism inspired by Kalman filtering. By utilizing neural networks to approximate unknown state equations, the system iteratively refines its internal estimates. This process allows the filter to adapt to changing dynamics without requiring explicit mathematical models of the environment.

The study implements both feedforward and recurrent neural networks to approximate system dynamics. These architectures serve as the primary tools for learning the underlying patterns of the nonlinear systems, allowing the filters to function effectively even when the initial process model is inaccurate.

The authors suggest that the two-step prediction-update structure is necessary to maintain consistency with established filtering theory. This framework allows the neural network to integrate new observations into the existing state estimate, ensuring that the filter remains responsive to incoming data streams.

Related Experiment Videos

Last Updated: Jun 30, 2026

Design and Application of a Fault Detection Method Based on Adaptive Filters and Rotational Speed Estimation for an Electro-Hydrostatic Actuator
06:45

Design and Application of a Fault Detection Method Based on Adaptive Filters and Rotational Speed Estimation for an Electro-Hydrostatic Actuator

Published on: October 28, 2022

Main Methods:

The review approach involves a comparative analysis of state estimation algorithms applied to nonlinear environments. Researchers implemented a two-step prediction-update framework to facilitate adaptive filtering. They utilized both feedforward and recurrent neural network architectures to approximate unknown system equations. The team conducted experiments using both off-line and online learning stages to evaluate filter performance. They developed extended Kalman filters as a baseline for measuring the efficacy of their neural-based models. The study design focused on testing these algorithms across various case studies with differing levels of dynamic complexity. Investigators assessed the convergence properties and estimation accuracy of each filter type under unknown noise statistics. This systematic evaluation allowed for a direct comparison between traditional mathematical techniques and the proposed machine learning-based approaches.

Main Results:

Key findings from the literature indicate that neural-based filters achieve convergence in complex scenarios where extended Kalman filters fail entirely. In one specific case study, the extended Kalman filter converged but produced higher state estimation errors than the neural counterparts. The authors report that off-line trained neural filters converge rapidly and exhibit acceptable performance levels. Online training further enhances the estimation accuracy of the developed adaptive filters. This training process effectively decouples the eventual filter accuracy from the initial accuracy of the process model. The research demonstrates that these algorithms make minimal assumptions regarding underlying nonlinear dynamics. The findings highlight the robustness of recurrent architectures in handling unpredictable noise statistics. The study confirms that adaptive learning significantly improves tracking precision compared to static model-based estimation methods.

Conclusions:

The authors demonstrate that neural-based filters offer a viable alternative to traditional estimation techniques in complex scenarios. Synthesis and implications suggest that these models handle unknown dynamics more effectively than standard mathematical approaches. The researchers report that neural filters maintain stability where extended Kalman filters often fail to converge. Their findings indicate that online training provides a significant boost to overall estimation precision. This approach effectively separates the final accuracy of the filter from the initial quality of the process model. The study confirms that recurrent architectures provide robust performance across diverse, challenging system configurations. These results highlight the potential for adaptive learning in environments where system rules are difficult to define. The authors conclude that their proposed algorithms represent a flexible solution for tracking hidden states in unpredictable dynamic systems.

Online training plays a vital role by continuously updating the neural network weights based on real-time data. This process decouples the final filter accuracy from the initial model quality, allowing the system to improve its performance as more information becomes available during operation.

The researchers measured state estimation errors to evaluate performance. In complex cases, the neural filters achieved convergence where extended Kalman filters failed, demonstrating superior stability and reduced error rates compared to the traditional mathematical benchmarks used in the study.

The authors propose that their adaptive approach provides a robust solution for systems with unknown dynamics. They imply that decoupling filter accuracy from model precision allows for more reliable state tracking in real-world applications where system rules are difficult to define or change over time.