LSTM Autoencoders for Internet of Things Data Compression and Battery Conservation

Abstract

This research focuses on a novel data compression technique in an Internet of Things (IoT) based digital communication system. The work simulated a wireless system being transmitted over a Rayleigh fading channel using Phase Shift Keying (M-PSK). The transmitted data has been sourced from Human Activity Recognition (HAR) using wearable devices and an open-source application. Unlike conventional compression techniques, this study uses Long Short-Term Memory (LSTM) Autoencoders to transform data in low-dimensional form in the wireless system. Primarily, the study evaluates the effectiveness of LSTM Autoencoders in data dimensionality reduction while preserving critical information for accurate activity recognition. This approach leads to better conservation in IoT devices. The proposed method's performance has been evaluated at various compression to the modulation levels. The Bit Error Rate (BER) vs. Signal Noise Ratio (SNR) curves of the M-PSK system have been evaluated in comparison with the Mean Square Error (MSE) of the compression and decompression. A trade-off between compression ratio and MSE has been illustrated, which ultimately leads to determining the accuracy of the measurement of human activity. The results highlight the benefits of leveraging the power of LSTM Autoencoders for data compression in the communication of the wireless system. The results contribute to advancing wearable HAR systems and this general performance optimization of the IoT system.