NeuCoReClass AD: Redefining self-supervised time series anomaly detection

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Time series anomaly detection is crucial for real-world applications.
• Unsupervised self-supervised learning models normal behavior without labeled data.
• Existing methods often use single proxy tasks and domain-specific transformations, limiting performance and generalization.

Purpose of the Study:

• Introduce NeuCoReClass AD, a self-supervised multi-task framework for time series anomaly detection.
• Address limitations of existing methods by employing multiple proxy tasks and neural transformation learning.
• Enable unsupervised characterization of anomaly profiles.

Main Methods:

• NeuCoReClass AD combines contrastive, reconstruction, and classification proxy tasks.
• Utilizes neural transformation learning for generating informative, diverse, and coherent augmented views.
• Avoids reliance on domain-specific handcrafted transformations.

Main Results:

• NeuCoReClass AD consistently outperforms classical and most deep-learning alternatives on various benchmarks.
• Demonstrates superior performance in unsupervised time series anomaly detection.
• Successfully characterizes distinct anomaly profiles in a fully unsupervised manner.

Conclusions:

• NeuCoReClass AD offers a robust and generalizable self-supervised approach for time series anomaly detection.
• The multi-task framework and neural transformations enhance the ability to model normal data patterns.
• This method advances the field by providing effective anomaly detection and characterization without labeled data.