Fault detection on manifolds of nonlinear dynamical systems with dual autoencoders

What the math gives to ML

The paper offers a transferable mechanism for converting nonlinear dynamical trajectories into structured Koopman-inspired delay-coordinate manifolds before applying autoencoding. Instead of reconstructing individual observations, it reconstructs coupled past and future Hankel blocks, so the learned latent variables must preserve temporal evolution and approximate a finite-dimensional Koopman shift. The most useful neural-network transfer is a dual sequence autoencoder with a statistically calibrated manifold-consistency test, applicable to anomaly detection, world models, and long-horizon sequence prediction.