Unsupervised Anomaly Detection for IoT-Based Multivariate Time Series: Existing Solutions, Performance Analysis and...

Area of Science:

• Computer Science
• Data Science
• Engineering

Background:

• Digitalization drives sensor deployment, generating vast unlabeled multivariate time series data.
• Multivariate Time Series Anomaly Detection (MTSAD) is crucial for identifying system anomalies but faces challenges in analyzing temporal and spatial dependencies.
• Unsupervised MTSAD is highly desirable due to the impracticality of labeling massive datasets.

Purpose of the Study:

• To provide a comprehensive review of the state-of-the-art in unsupervised Multivariate Time Series Anomaly Detection (MTSAD).
• To offer a theoretical background on MTSAD techniques.
• To present a numerical evaluation of prominent unsupervised MTSAD algorithms.

Main Methods:

• Review of advanced machine learning, signal processing, and deep learning techniques for unsupervised MTSAD.
• Theoretical exposition of MTSAD principles, emphasizing temporal and spatial dependency analysis.
• Comparative numerical evaluation of 13 selected unsupervised MTSAD algorithms using two public datasets.

Main Results:

• Identification of key challenges and advancements in unsupervised MTSAD.
• Performance comparison of 13 algorithms, highlighting their strengths and weaknesses.
• Empirical evidence on the effectiveness of various unsupervised approaches for MTSAD.

Conclusions:

• Unsupervised MTSAD methods are critical for handling large-scale, unlabeled sensor data in industrial and other applications.
• The reviewed algorithms offer diverse capabilities for detecting anomalies in multivariate time series.
• Further research is needed to address the complexities of simultaneous temporal and spatial dependency analysis in MTSAD.