Ensemble Neuroevolution-Based Approach for Multivariate Time Series Anomaly Detection.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Multivariate time series anomaly detection is critical for industrial failure prevention, but human analysis is difficult due to large sensor data.
• Existing automated methods often rely on autoencoder and generative adversarial network architectures.
• The need for efficient and automated anomaly detection is growing with industrial system complexity.

Purpose of the Study:

• To present a novel framework using neuroevolution to automatically enhance anomaly detection models.
• To improve the performance of both new and existing deep learning anomaly detection models.
• To demonstrate the effectiveness of an automated approach for building ensemble anomaly detection systems.

Main Methods:

• The framework employs evolution strategies to develop an ensemble model, where each component model analyzes a subset of sensor data.
• Neuroevolution is utilized to optimize model architecture and hyperparameters, including window size, number of layers, and layer depths.
• The approach was tested on the SWAT and WADI datasets for anomaly detection.

Main Results:

• The proposed neuroevolution framework successfully boosted the anomaly detection scores of various deep learning models.
• The system operates in a fully automated mode, optimizing ensemble models within a reasonable timeframe.
• The study achieved significant performance improvements on benchmark datasets.

Conclusions:

• Neuroevolution offers a powerful and automated method for optimizing ensemble deep learning models for anomaly detection.
• This is the first known approach to automatically construct an ensemble deep learning anomaly detection model using neuroevolution.
• The framework provides a scalable and efficient solution for enhancing failure prevention systems in industries.