Enhanced LSTM autoencoder framework for robust anomaly detection in sensor networks
|
Full Text |
Pdf
|
|
Author |
Abbas Ibrahim Al-Zaki
|
|
e-ISSN |
1819-6608 |
|
On Pages
|
196-209
|
|
Volume No. |
21
|
|
Issue No. |
3
|
|
Issue Date |
April 10, 2026
|
|
DOI |
https://doi.org/10.59018/022630
|
|
Keywords |
anomaly detection, LSTM autoencoder, ELSTM-AE, EEG, sensor networks.
|
Abstract
Anomaly detection in modern industrial and Internet of Things (IoT) sensor networks is critically challenging due to the voluminous, multivariate time-series data, which is often contaminated by noise, missing values, cyber-attacks, or operational malfunctions. Traditional statistical methods and basic artificial intelligence models prove insufficient for capturing the intrinsic nonlinear temporal dependencies within this complex data. We propose a new network architecture called the Enhanced Long Short-Term Memory Autoencoder (ELSTM-AE) to overcome these problems. It is meant to be very good at finding anomalies in sensor networks. The ELSTM-AE has a complex structure that includes multi-stage LSTM encoding and decoding, batch normalization, dropout regularization, and adaptive normalization. This makes the model more accurate and stronger. We conducted thorough experimentation using real-world industrial testbeds, namely the Swat and WADI datasets. The results demonstrate very promising performance improvements, with the ELSTM-AE achieving an accuracy of up to 94.48%, a significant increase in recall to 88.06% (from 34.09%), and an F1-score of 0.8811 after optimal threshold tuning. Further analyses, encompassing sensitivity and Friedman tests, validated that our model significantly surpasses traditional baselines such as Z-score and moving average. The proposed ELSTM-AE model provides an easy-to-understand and scalable way to find anomalies in real time in important cyber-physical systems.
Back