Towards Outlier Sensor Detection in Ambient Intelligent Platforms-A Low-Complexity Statistical Approach.

Area of Science:

• Computer Science
• Data Science
• Network Engineering

Background:

• Sensor networks in smart cities generate vast, heterogeneous data streams.
• Detecting outliers is critical for maintaining secure and reliable databases in these environments.
• Existing outlier detection methods often require significant computational and storage resources, unsuitable for sensor nodes.

Purpose of the Study:

• To analyze and compare three statistical prediction models for outlier detection in sensor networks.
• To identify a low-complexity model that minimizes memory consumption and computational time.
• To provide validated datasets for future research in real-world outlier detection.

Main Methods:

• Evaluation of three binary classifiers based on Auto-Regressive Integrated Moving Average (ARIMA), Generalized Additive Model (GAM), and LOcal RegrESSion (LOESS) models.
• Focus on models with low memory footprint and computational time suitable for resource-constrained sensor nodes.
• Development of real-world classified datasets to serve as ground truth.

Main Results:

• The Auto-Regressive Integrated Moving Average (ARIMA) model demonstrated superior performance as a classifier for outlier detection.
• The study identified optimal settings for the ARIMA model for effective outlier identification.
• Two real-world datasets, classified for outlier detection, were generated.

Conclusions:

• The ARIMA model is the recommended approach for efficient outlier detection in sensor network environments.
• The developed datasets offer valuable resources for advancing research in real-world outlier detection and data reliability.
• The findings contribute to the development of more robust and efficient data management strategies for smart city applications.