Big Data Dimensionality Reduction for Wireless Sensor Networks Using Stacked Autoencoders

Abstract

A typical wireless sensor network (WSN) consists of spatially dispersed sensor nodes for monitoring and gathering data at a central location. Over time, this data accumulates and it becomes difficult to transmit this large volume of data. Owing to the recent advancements in internet of things (IoTs), more complex data is generated. This high-dimensional data must be reduced to a lower dimensional representation before transmission. Principal component analysis (PCA) is a common method for data aggregation. However, PCA performs best for the cases that require linear mapping. Therefore, a method for linear and/or non-linear data mapping is considered in this work. Autoencoders are neural networks that can learn about the internal structure of the data and reduce its dimensionality. Stacked autoencoders are a class of deep feed forward neural networks that perform real learning of the data. Principal component analysis assumes that the underlying data can be fully described by its mean and variance, whereas, autoencoders are adaptive to different data distributions. In this paper, we compare the performance of PCA and autoencoder in terms of their respective ability to reduce the reconstruction error in the observed high-dimensional data from a WSN. The results of numerical analysis demonstrate that the autoencoder reduces the reconstruction error by up to 10 times as compared to an ordinary PCA-based scheme.