Anomaly Detection Using an Ensemble of Multi-Point LSTMs.

Area of Science:

• Data Science
• Artificial Intelligence
• Machine Learning

Background:

• The proliferation of time-series data from sources like smartwatches and smart factories necessitates robust anomaly detection.
• Traditional univariate time-series anomaly detection methods are insufficient for complex, modern datasets.
• Deep learning approaches are increasingly vital for accurate time-series anomaly detection in various industries.

Purpose of the Study:

• To propose a novel deep learning-based anomaly detection algorithm for time-series data.
• To develop an ensemble of multi-point Long Short-Term Memory (LSTM) networks adaptable to multiple time-series domains.
• To enhance the accuracy and efficiency of detecting abnormal patterns in complex time-series data.

Main Methods:

• A three-step approach involving automatic model selection, ensemble stacking of multiple LSTMs, and final anomaly detection.
• Utilizing an ensemble of multi-point LSTMs to process and analyze time-series data.
• Comparing the proposed model against state-of-the-art deep learning models on three real-world datasets.

Main Results:

• The proposed ensemble LSTM model achieved excellent accuracy and a good F1 score across three diverse datasets.
• Demonstrated efficient execution time compared to other advanced time-series anomaly detection models.
• The model effectively detects anomalies in complex, multivariate time-series data.

Conclusions:

• The ensemble of multi-point LSTMs offers a powerful and versatile solution for time-series anomaly detection.
• The proposed method provides a significant advancement in detecting cyber-intrusions, fraud, and industrial anomalies.
• While training time is a consideration, the model's performance justifies its use in critical applications.