Skip to content
Conference Open access

Comparative Analysis of LSTM and Convolutional Autoencoders for Root-Cause Localization in Multivariate Time-Series Anomalies

Sep 2026 · 2026 11th International Conference on Computer Science and Engineering (UBMK) · pp. 918-923 · 0 citations · 24 references

Abstract

In multivariate time-series monitoring systems, merely detecting an anomaly is often insufficient; effective intervention also requires identifying which measurement dimension (sensor or component) is responsible for it. From a test-and-verification engineering perspective, this study addresses anomaly detection and root-cause localization within a single unsupervised framework and compares two reconstruction-based models under identical conditions: a recurrent Long Short-Term Memory Autoencoder (LSTM-AE) and a one-dimensional Convolutional Autoencoder (Conv-AE). Both models are processed through the same preprocessing, windowing, Peaks-Over-Threshold (POT) thresholding, and point-adjusted (PA) evaluation pipeline. To identify the responsible dimensions, a reconstruction-residual-based, dimension-level attribution method applicable to any reconstruction-based model without modification is employed. Experiments on the NASA (National Aeronautics and Space Administration) SMAP (Soil Moisture Active Passive) and MSL (Mars Science Laboratory) telemetry datasets and the Server Machine Dataset (SMD) show that both models achieve comparable detection performance across the three datasets; however, LSTM-AE is markedly superior in root-cause localization on SMD, which is specifically used for quantitative evaluation due to the availability of dimension-level ground-truth labels. The results indicate that strong detection performance does not guarantee strong diagnostic performance, and that root-cause localization should therefore be treated as an independent and critical capability in monitoring systems.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.