Reconstituição temporal de dados na internet das coisas usando redes neurais: uma comparação com métodos clássicos de interpolação

Abstract

Currently, we live in the era of Big Data, in which cutting-edge scientific instruments, networks, social media, as well as Internet of Things (IoT) devices generate and transmit enormous amounts of data daily through the internet. In the context of IoT, a new paradigm emerges to mitigate the overload in sending such data to the cloud. Thus, in contrast to cloud computing, fog computing is increasingly being adopted, with most of the data processing and storage being performed at the network edge by gateways. In this scenario, in environments where sensors collect information every second, for example, instead of sending each measurement to the cloud, the gateway adopts a data aggregation strategy before transmitting them. However, since various applications require data as originally measured by the devices, and considering that the cloud has greater capacity to serve a larger number of clients compared to gateways, the ideal scenario implies the cloud's ability to reconstruct the data history and provide it to these applications. This work proposes to investigate the capacity of the perceptron neural network to reconstruct sensor data history through interpolation from the aggregated values sent to the cloud. Since IoT devices are resource-constrained, the study also compares the perceptron neural network developed with classical interpolation algorithms to evaluate the efficiency and effectiveness between the methods. To achieve this goal, two neural networks were developed: the first at the network edge, which learns the behavior of sensor data over time, while the second neural network located in the cloud performs interpolation based on the aggregated data sent and the model generated by the edge neural network. The results obtained indicate that even a relatively simple neural network architecture like the perceptron can perform IoT data interpolation with a considerable margin of accuracy. When compared to classical interpolation algorithms, considering criteria such as time and Mean Squared Error (MSE), the perceptron network proves to be statistically as effective as classical methods but less efficient in interpolation time. In summary, this study contributes to understanding the possibilities and limitations of applying neural networks in the temporal reconstruction of IoT data and highlights the need to evaluate the same comparison with the use of other neural network architectures, such as Long Short-Term Memory (LSTM) neural networks.