PASE: An autonomous sequential framework for the state estimation of dynamical systems

Kalman filter (KF) and its variants are commonly used in estimating states of dynamical systems. For accurate estimation of states, researchers have developed intelligent KFs by combining them with machine learning algorithms. The inherent variations in the vehicle dynamics and uncertainties in the environment require the development of an autonomous sequential framework for continuous estimation. Further, the need for real-time estimation of states in certain applications requires low computer memory usage and low computational cost while implementing autonomous-structured learning algorithms. In this paper, a parsimonious autonomous sequential estimator (PASE) is proposed, which combines the KF-based estimator and autonomous-structured recurrent parsimonious learning machine (rPALM) in a sequential manner. The rPALM overcomes the dependency on target variables while point-to-hyperplane distance calculation. The performance of PASE has been evaluated extensively by comparing it with various batch-learning algorithms and single-pass learning-based intelligent estimators. The results clearly indicate that PASE provides better estimation accuracy with a compact architecture for both linear and nonlinear dynamical systems. Finally, the performance of PASE has been evaluated with experimental data for the state estimation of an unmanned ground vehicle while the training of the learning machine is performed with the simulated data. The estimation accuracy in such a scenario is justifying its appropriateness in real-world applications.

Automated summary

Researchers have developed intelligent Kalman filters (KFs) by combining them with machine learning algorithms for accurate estimation of states in dynamical systems. In this paper, a parsimonious autonomous sequential estimator (PASE) is proposed, which combines a KF-based estimator and an autonomous-structured recurrent parsimonious learning machine (rPALM) in a sequential manner. The performance evaluation shows that PASE provides better estimation accuracy with a compact architecture for both linear and nonlinear dynamical systems, making it suitable for real-world applications.