An enhanced Smart Sampling algorithm based on Deep Learning

The spread of Wireless Sensor Networks, driven mainly by the increasing use of pervasive sensors in everyday reality, according to the IoT concept, is bringing to light increasing issues in terms of energy consumption and bandwidth occupancies for the transmission of acquired data. In this regard, researchers and standards committees, especially that of the IEEE 21451 standard, have focused their efforts on Smart Sampling methods to reduce the number of samples acquired and then transmitted over the network adaptively based on the dynamics of the measured signal. This work aims to employ new techniques based on deep learning, especially LSTM autoencoders, to predict certain signal time windows to turn off the acquisition and transmission system, delegating the full signal computation to a central processing unit. Thus, the objective was to evaluate the accuracy and timeliness of the method to replace the classical Real-Time Segmentation algorithm proposed by the IEEE 21451 standard. The results obtained on three types of signals were satisfactory, clearing the way for implementation in smart data acquisition systems