Adaptive Prototypical Contrastive Learning for Time Series Clustering

Unsupervised analysis of time series is pivotal for IoT, healthcare, and industrial monitoring, yet standard methods face a fundamental dilemma: representation learning relies on instance discrimination, while clustering requires semantic grouping. Worse still, existing deep clustering methods depend on a pre-defined cluster count K, a parameter rarely known in real-world streams. We propose Adaptive Prototypical Contrastive Learning (APCL), a self-evolving framework that jointly learns representations and discovers cluster structure. APCL introduces (1) MDL-guided Prototype Evolution to dynamically split/merge centers based on a geometric consistency criterion, (2) Soft-to-Hard Annealing to handle ambiguous boundaries, and (3) Hierarchical Consistency to capture multi-scale patterns. APCL transforms the ''guesswork'' of K-selection into an optimization problem. On complex non-linear datasets like HAR, where fixed-K baselines are sensitive to misspecification, APCL automatically identifies K^*, achieving substantial ARI improvement. Critically, we demonstrate that APCL maintains robustness even on simpler, linearly-separable data where other deep methods struggle against PCA. Across five diverse domains, APCL achieves strong performance on complex, non-linear datasets and remains highly competitive on simpler tasks compared to methods equipped with the ground-truth K. Crucially, APCL helps bridge the gap between representation learning and structure discovery, showing that robust unsupervised learning is achievable without prior knowledge of cluster counts.