Hierarchical ODE Learns Continuous Prototypes for Early Failure Detection.

Learning prototypes from time series data is challenging due to observational ambiguity, where discrete architectures struggle to separate stochastic noise from continuous dynamics. Furthermore, rigid assumptions about a fixed number of prototypes often fail to capture the full diversity of unseen data. To overcome these limitations, a new hierarchical ordinary differential equation (ODE) clustering network is introduced. This network uses neural ODEs to model latent state evolution as a continuous integral curve, which enforces temporal continuity. This approach effectively disentangles smooth feature trends from random noise. The adaptive hierarchical mechanism within the network autonomously determines the appropriate number of prototypes without requiring prior constraints. The method was validated on the task of early link failure detection using irregularly sampled time series, demonstrating its ability to extract underlying physical prototypes and achieve robust failure detection.