Anomaly Detection in Multivariate Time Series with Contaminated Training Data Using VAE

• Published in: 2024 IEEE 10th Conference on Big Data Security on Cloud (BigDataSecurity) pp. 5 - 10
• Main Authors: Lv, Dongfeng, Ma, Manman, Dong, Hanxia, Shang, Yishan, Bao, Kuo
• Format: Conference Proceeding
• Language: English
• Published: IEEE 10.05.2024
• Subjects: anomaly detection; Data models; Deep learning; multivariate time series; Noise measurement; noisy data; pseudo label; Security; Time series analysis; Training; Training data; variational autoencoder
• DOI: 10.1109/BigDataSecurity62737.2024.00010

Multivariate time series anomaly detection (MVT-SAD) is a significant data mining task with numerous applications in the IoT era. Nowadays, with the advancement of deep learning technology, an increasing number of deep learning models have made significant progress in the MVTSAD field compared to traditional models. Unsupervised deep learning models learn the latent representation of normal data and then detect anomalies based on reconstruction error. However, as the volume of MVTS data increases, and the training set consists entirely of normal samples, which is often impractical to achieve, accurately capturing the distribution of normal data and subsequently detecting anomalies becomes challenging. In this work, to address this crucial challenge, we propose a Variational Autoencoder (VAE)-based anomaly detection method for MVTS. The core idea is to establish pseudo-labels during the training process to identify noise values in the training set and establish a training objective to eliminate training errors, thereby obtaining a robust model learned from noisy data. Experimental results reveal that our method effectively mitigates performance degradation in mainstream deep MVTSAD models, achieving at least a 1.25% increase in F1 score.