Deep Learning Techniques In Predicting Average Localization Error In Wireless Sensor Networks (WSNs)

Abstract

WSNs require accurate localization, but the measurement errors, such as
noise, environmental conditions, and network dynamics, result in a high ALE.
Despite the many methods that have been proposed to improve localization
error, the prediction of ALE has not yet gained sufficient attention. In this
paper, we propose the first deep learning model to predict the ALE of WSNs
as a supervised regression problem. We create a thoroughly synthetic dataset
through experiments over different network parameters such as network
density, anchor ratio, communication range, and noise models. We develop a
hybrid deep learning model, CNN, GCN, and LSTM, that takes advantage of
all kinds of spatial, topological, and temporal biases of localization error.
Comprehensive evaluations indicate that the deep learning models
dramatically outperform the traditional machine learning baselines (R = 0.94),
and the CNN achieved the highest performance (R = 0.982, MAPE = 5.48%).
The hybrid model, combining CNN, GCN, and LSTM, is able to obtain the
expected performance (R = 0.964, MAPE = 8.63%) and sets a reachable
baseline for ALE prediction. We highlight that the complexity of GCN and
stacking does not provide an extra improvement in performance on this dataset and guides us in designing the deep learning models for future work. This
paper proposes the first deep learning baseline to estimate the ALE, which
assists the network designers in predicting the localization performance and
optimally allocating resources for the network deployment. Source code and
dataset can be downloaded at: https://github.com/DataCodepro/ALEWNS

Journal Documents