MTS-JEPA: Multi-Resolution Joint-Embedding Predictive Architecture for Time-Series Anomaly Prediction

TL;DR

The paper addresses proactive anomaly prediction in multivariate time series by identifying two weaknesses of single-resolution JEPA: representation collapse and an inability to capture precursor signals across scales. It introduces MTS-JEPA, a multi-resolution JEPA with a soft codebook bottleneck and a dual-view teacher–student framework that predicts future latent states at multiple temporal scales. The approach yields theoretical stability insights, including upper bounds on latent drift and non-collapse guarantees, and demonstrates state-of-the-art performance on four benchmark datasets with robust ablations and regime visualizations. This method offers a scalable, interpretable foundation for early warning in critical infrastructure, with improved resilience to distribution shifts due to its discrete, regime-focused latent dynamics.

Abstract

Multivariate time series underpin modern critical infrastructure, making the prediction of anomalies a vital necessity for proactive risk mitigation. While Joint-Embedding Predictive Architectures (JEPA) offer a promising framework for modeling the latent evolution of these systems, their application is hindered by representation collapse and an inability to capture precursor signals across varying temporal scales. To address these limitations, we propose MTS-JEPA, a specialized architecture that integrates a multi-resolution predictive objective with a soft codebook bottleneck. This design explicitly decouples transient shocks from long-term trends, and utilizes the codebook to capture discrete regime transitions. Notably, we find this constraint also acts as an intrinsic regularizer to ensure optimization stability. Empirical evaluations on standard benchmarks confirm that our approach effectively prevents degenerate solutions and achieves state-of-the-art performance under the early-warning protocol.

Figures (8)

• Figure 1: Comparison between standard single-resolution JEPA (A) and the proposed multi-resolution JEPA (B).
• Figure 2: Overview of the MTS-JEPA framework for time series representation learning.
• Figure 3: Window-level comparisons of dominant latent codes between anomalous and normal windows on the PSM dataset. Shaded regions indicate patch-level segments where the dominant code is selected.
• Figure 4: Histogram comparing the occurrence probability of latent codes in anomalous versus normal windows on the PSM test set. The codes are sorted by their anomaly–normal probability difference (from largest to smallest). PSM is used because it provides a large test set with reliable window-level anomaly labels.
• Figure 5: Overview of the Proposed Framework.

Theorems & Definitions (12)

• Definition 1.1: Soft Code Map
• Lemma 1.2: $\ell_1 \to \ell_2$ Lipschitzness
• proof
• Theorem 1.3: Stability Bound
• proof
• Remark 1.6
• Lemma 1.8: Entropy lower bound implies at least two active codes
• proof
• Lemma 1.9: Marginal mass transfers to dominant-code frequency
• proof