Abhishek Bhardwaj , Max Quinn and Apeksha Agnihotri , Amazon Web Services, USA

This paper presents SALT (Sustained Anomaly Labeling in Time Series), a novel approach combining autoencoder neural networks with gradient-based filtering for detecting sustained anomalies in time series data. While traditional anomaly detection methods often struggle with distinguishing between transient spikes and meaningful sustained anomalies, SALT addresses this limitation by leveraging autoencoders ability to learn normal data pat terns and incorporating gradient filtering to identify persistent deviations. The method was evaluated on both curated data and real-world datasets from the Numenta Anomaly Bench mark (NAB). Results demonstrate significant improvements in precision while maintaining high recall- achieving F1 scores of 0.91 and 0.86 on curated and real-world data respectively, compared to 0.72 and 0.48 using autoencoders alone. SALT successfully reduces false posi tives from transient anomalies while accurately detecting sustained anomalous patterns. The framework shows promise for applications in network security, industrial systems monitoring, and other domains requiring reliable detection of persistent anomalies in streaming data.

Sustained Anomaly, Gradient Filtering, Deep Learning, Data Labeling, Time Series, Autoencoder