Time Series Anomaly Detection Model Based on Multi-Features.

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Monitoring key time series data is crucial for anomaly detection in internet information services.
• Existing anomaly detection models face challenges in practical application due to repeated parameter tuning and the need for model selection for different data types.

Purpose of the Study:

• To propose an automated and adaptable anomaly detection model for time series data.
• To address the practical limitations of current anomaly detection methods in real-world internet service monitoring.

Main Methods:

• Feature engineering including statistical, fitting, and time-frequency domain features for time series.
• Utilizing a random forest ensemble model for automated feature selection and anomaly classification.
• Introducing the Anomaly Detection Capability (ADC) score, incorporating a timeliness window to account for detection delay, building upon the F1-score.

Main Results:

• The proposed model demonstrates effective anomaly detection capabilities on Key Performance Indicator (KPI) time series data.
• The ADC score achieved by the model ranges from 0.7 to 0.8, indicating practical applicability.
• The automated feature selection and classification approach reduces the need for manual parameter adjustment.

Conclusions:

• The developed time series anomaly detection model offers a practical solution for monitoring internet information services.
• The ADC score provides a more comprehensive evaluation of anomaly detection performance by including timeliness.
• The model's adaptability and automated feature selection make it suitable for diverse time series data encountered in industry.