Enhancing Time Series Anomaly Detection: A Knowledge Distillation Approach with Image Transformation.

Area of Science:

• Data Science
• Machine Learning
• Computer Vision

Background:

• Anomaly detection is vital in safety-critical domains but hindered by limited abnormal data and high labeling costs.
• Time series anomaly detection presents unique challenges due to sequential data, computational demands, and noise.
• Image anomaly detection has seen significant advancements, offering high accuracy and efficiency.

Purpose of the Study:

• To develop a novel framework for time series anomaly detection by integrating image-based techniques.
• To leverage Gramian Angular Field (GAF) transformations for converting time series data into images.
• To apply advanced image anomaly detection models, Reverse Distillation (RD) and EfficientAD (EAD), to transformed time series data.

Main Methods:

• Time series data converted to images using Gramian Angular Field (GAF) transformations.
• Application of state-of-the-art image anomaly detection models: Reverse Distillation (RD) and EfficientAD (EAD).
• Implementation of tailored preprocessing and transformation techniques for enhanced performance and interoperability.

Main Results:

• The proposed framework demonstrated high overall recall across various datasets.
• Achieved approximately 99% F1 scores on specific univariate datasets, indicating high accuracy.
• Successfully applied image anomaly detection techniques to time series data, proving framework effectiveness.

Conclusions:

• The novel framework effectively bridges image anomaly detection with time series analysis.
• The GAF transformation combined with RD and EAD offers an efficient and accurate solution for time series anomaly detection.
• This approach addresses key challenges in time series anomaly detection, including data scarcity and computational costs.