Data-Driven Anomaly Detection Approach for Time-Series Streaming Data.

Area of Science:

• Computer Science
• Electrical Engineering
• Environmental Science

Background:

• Wireless sensor networks (WSNs) are crucial for environmental monitoring but face challenges with sensor node failures.
• Fault diagnosis in WSNs is critical for maintaining network reliability and stability, especially with large-scale deployments.
• Time-series streaming data from sensor nodes requires robust anomaly detection methods.

Purpose of the Study:

• To propose and evaluate a data-driven anomaly detection approach for time-series status data from WSN sensor nodes.
• To enhance the accuracy and efficiency of fault diagnosis in WSNs operating in harsh environments.
• To improve the reliability and stability of WSNs through effective anomaly detection.

Main Methods:

• A novel approach combining a median filter (MF) for preprocessing, stacked long short-term memory (LSTM) for prediction, and exponentially weighted moving average (EWMA) control charts for anomaly detection.
• Preprocessing of raw sensor data, including operating voltage and panel temperature, using MF to handle obvious anomalies.
• Utilizing stacked LSTM for time-series prediction and EWMA for anomaly recognition in the processed data.

Main Results:

• The proposed MF-stacked LSTM-EWMA approach demonstrated significant improvements in detection rate (DR) and false rate (FR) compared to other methods.
• Achieved an average DR of 95.46% and an average FR of 4.42% on real-world time-series status data.
• The MF-stacked LSTM-EWMA method obtained a superior F2 score, indicating better overall performance in anomaly detection.

Conclusions:

• The MF-stacked LSTM-EWMA approach is effective for anomaly detection in WSNs, particularly for time-series status data from sensor nodes in harsh conditions.
• This data-driven method offers valuable insights for improving fault diagnosis and enhancing the reliability of WSNs.
• The approach provides a robust solution for identifying anomalies in environmental monitoring WSNs.