Self-Supervised Time-Series Anomaly Detection Using Learnable Data Augmentation

TL;DR

LATAD tackles the challenge of anomaly detection in multivariate time series under limited labeled data by introducing a self-supervised framework that uses learnable data augmentation and triplet-based contrastive learning to generate highly discriminative latent representations. The method fuses inter-feature correlations and temporal context via a combined 1-D convolution, GAT, transformer encoder, and TCN, producing latent vectors $\mathbf{z}$ used in a triplet objective with positives from a temporal neighborhood and negatives generated by learnable masks. Training optimizes $\mathcal{L}_{\mathrm{LATAD}}=\mathcal{L}_{\mathrm{comp}}+\mathcal{L}_{\mathrm{sep}}+\lambda\mathcal{L}_{\mathrm{reg}}$, and anomalies are scored in latent space with a K-means-based metric $\mathcal{A}(\mathbf{x})$, enabling robust detection and gradient-based root-cause interpretation. Across five real-world datasets, LATAD achieves comparable or superior F1 metrics to state-of-the-art methods, with notable gains on MSL, SMAP, and SMD, and demonstrates practical anomaly diagnosis via gradient analysis, supporting industrial deployment under data scarcity.

Abstract

Continuous efforts are being made to advance anomaly detection in various manufacturing processes to increase the productivity and safety of industrial sites. Deep learning replaced rule-based methods and recently emerged as a promising method for anomaly detection in diverse industries. However, in the real world, the scarcity of abnormal data and difficulties in obtaining labeled data create limitations in the training of detection models. In this study, we addressed these shortcomings by proposing a learnable data augmentation-based time-series anomaly detection (LATAD) technique that is trained in a self-supervised manner. LATAD extracts discriminative features from time-series data through contrastive learning. At the same time, learnable data augmentation produces challenging negative samples to enhance learning efficiency. We measured anomaly scores of the proposed technique based on latent feature similarities. As per the results, LATAD exhibited comparable or improved performance to the state-of-the-art anomaly detection assessments on several benchmark datasets and provided a gradient-based diagnosis technique to help identify root causes.

Figures (7)

• Figure 1: Overall architecture of LATAD.
• Figure 2: Illustration of how the two attention-based modules to model correlations and temporal dependencies in the multivariate time-series data. (a) Graph attention module. (b) Transformer encoder module.
• Figure 3: Triplet-based contrastive learning framework with positive samples and negative samples. (a) Positive samples from temporal neighborhood. (b) Generator-produced negative samples for the given input.
• Figure 4: Case study results of anomaly diagnosis. (a) Top $k$ root causes. (b) Trend charts of Top $k$ root causes on December 31, 2015, at 01:45:19.
• Figure 5: Visualization of extracted feature representations for different methods (best viewed in color). (a) MTAD-GAT and (b) LATAD for the SWaT dataset. The green points are from normal class and red one are from anomalous class.