Self-Supervised Learning for Interpretable Time Series Representations

A novel technique for deriving interpretable representations from progressive time series has been presented in a paper on arXiv (2605.31061). This method employs a self-supervised contrastive objective to establish a low-dimensional latent space, where each data point corresponds to a location on a manifold defined by two fixed orthogonal prototype vectors. The resulting trajectories traverse this manifold, uncovering a latent compass characterized by polar coordinates (θ, r): θ indicates progression (e.g., from healthy to failed), while r denotes the active mode (e.g., operating condition), all without using proxy labels. The approach was tested against leading methods in industrial degradation, robotic tasks, and neural activity, confirming three critical aspects.