Segmentation-Based Adversarial Denoising Auto-Encoder for Anomaly Detection in Multivariate Time Series Data

Abnormal detection is crucial for maintaining critical processes' stability, security, and efficiency in various domains. Traditional methods for abnormal detection in time series data have limitations in capturing temporal dependencies and providing interpretable anomaly scores. The authors propose a Segmentation-based Adversarial Denoising Auto-encoder (SADA) model to address these challenges. The SADA model uses two individual auto-encoders to establish an adversarial training process and incorporates segmentation techniques and inter-correlation encoding units to capture temporal dependencies and inter-correlations in multivariate time series data. Experimental results on multiple datasets demonstrate the effectiveness of the SADA model in detecting anomalies and achieving high precision and recall rates. The proposed model shows promise in improving the adaptability of anomaly detection algorithms to different types of data.

I. Introduction

Abnormal detection is of utmost importance in various domains and industries as it enables proactive monitoring, early detection of anomalies, and timely response to potential issues. [1] By analyzing data and identifying deviations from normal behavior, abnormal detection systems play a crucial role in maintaining critical processes' stability, security, and efficiency.