Anomaly Detection in Multivariate Time Series: Combining LSTM Autoencoders with Contrastive Learning

Abstract

We propose a Long Short-Term Memory(LSTM)-based autoencoder model for multivariate time series anomaly detection that incorporates contrastive learning tailored to time-series characteristics. By leveraging contrastive representation learning, the model effectively pulls normal data closer to the original representation while pushing anomalous data further away, enhancing detection performance. To generate positive and negative pairs, the model applies time series-specific augmentations by sampling overlapping segments, preserving contextual integrity. It combines both instance and temporal contrastive learning to capture richer representations. Training is guided by a joint loss function that integrates weighted contrastive loss with reconstruction loss. Experimental results demonstrate that the proposed method improves F1-score by 2–6% over baseline models. This work highlights that even with a simple LSTM-based autoencoder architecture, significant gains in anomaly detection can be achieved by incorporating contrastive learning strategies suited for time series data.