Ensemble Denoising Autoencoders Based on Broad Learning System for Time-Series Anomaly Detection.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Unsupervised time-series anomaly detection faces challenges due to scarce labels and complex anomaly definitions.
• Real-time detection requires low computational cost and high model robustness, which are often unmet.

Purpose of the Study:

• To enhance the recognition of anomalous patterns in unsupervised scenarios.
• To develop an efficient and robust unsupervised time-series anomaly detection method.

Main Methods:

• Proposing a data-driven spontaneous perturbation and sequence-image strategy.
• Introducing temporal anomaly knowledge enhancement using artificial anomalous data pairs.
• Developing denoising autoencoders based on the broad learning system (DBLS-AE) and progressively diverse denoising autoencoders (PddBLS-AE).

Main Results:

• PddBLS-AE effectively learns anomalous patterns for efficient anomaly detection.
• The model demonstrates improved robustness in handling diverse temporal anomalies.
• Accelerated training is achieved compared to advanced deep learning models using the broad learning system (BLS).

Conclusions:

• PddBLS-AE offers a robust and efficient solution for unsupervised time-series anomaly detection.
• The proposed methods significantly improve performance and robustness across multiple datasets.
• The broad learning system integration enables faster training and better anomaly cognition.