Abstract:Determination of the nature of the dynamical state of a system as a function of its parameters is an important problem in the study of dynamical systems. This problem becomes harder in experimental systems where the obtained data is inadequate (low-res) or has missing values. Recent developments in the field of topological data analysis have given a powerful methodology, viz. persistent homology, that is particularly suited for the study of dynamical systems. Earlier studies have mapped the dynamical features with the topological features of some systems. However, these mappings between the dynamical features and the topological features are notional and inadequate for accurate classification on two counts. First, the methodologies employed by the earlier studies heavily relied on human validation and intervention. Second, this mapping done on the chaotic dynamical regime makes little sense because essentially the topological summaries in this regime are too noisy to extract meaningful features from it. In this paper, we employ Machine Learning (ML) assisted methodology to minimize the human intervention and validation of extracting the topological summaries from the dynamical states of systems. Further, we employ a metric that counts in the noisy topological summaries, which are normally discarded, to characterize the state of the dynamical system as periodic or chaotic. This is surprisingly different from the conventional methodologies wherein only the persisting (long-lived) topological features are taken into consideration while the noisy (short-lived) topological features are neglected. We have demonstrated our ML-assisted method on well-known systems such as the Lorentz, Duffing, and Jerk systems. And we expect that our methodology will be of utility in characterizing other dynamical systems including experimental systems that are constrained with limited data.

Topology Learning in Radial Dynamical Systems with Unreliable Data

Topology Learning of unknown Networked Linear Dynamical System excited by Cyclostationary inputs

Topology Reconstruction of Heterogeneous Networked Dynamical Systems with Unknown Input Matrices

Topology Identification under Spatially Correlated Noise

Causal Structure Identification from Corrupt Data-Streams

Learning the Topology and Behavior of Discrete Dynamical Systems

Topology Recoverability Prediction for Ad-Hoc Robot Networks: A Data-Driven Fault-Tolerant Approach

Learning Topology of the Power Distribution Grid with and without Missing Data

Inferring network topology from complex dynamics

Online Topology Inference from Streaming Stationary Graph Signals with Partial Connectivity Information

Let's do the time-warp-attend: Learning topological invariants of dynamical systems

Learning Topology of Distribution Grids using only Terminal Node Measurements

Dual-based Online Learning of Dynamic Network Topologies

Online Joint Topology Identification and Signal Estimation from Streams with Missing Data

Topology Learning for Heterogeneous Decentralized Federated Learning Over Unreliable D2D Networks

Data-Driven Adaptive Consensus Learning From Network Topologies

Topology Estimation in Bulk Power Grids: Guarantees on Exact Recovery

Tractable learning in under-excited power grids

Network Topology Inference from Smooth Signals Under Partial Observability

Characterization of dynamical systems with scanty data using Persistent Homology and Machine Learning