Article Preview
Top1. Introduction And Motivation
Smart cities, Industrial 4.0, smart health monitoring, and more other internet of things applications are networks that integrate wireless sensor nodes to gather data (Lin et al., 2020)(Messai & Seba, 2016). Each sensor node could be able to sense different types of data such as humidity, pressure, temperature, vibration, pollution, etc. These many types of data collected make sensor nodes widely used in divers wireless sensor networks (WSNs) applications. A WSN is a distributed system that is generally composed of sensor nodes that have some characteristics. They have limited resources in terms of energy (battery-powered device), computation, and storage. Once deployed in the sensed area, these nodes sense data and collaborate to root them to a centralized special node called base station or sink for processing and decision support. Once deployed in the interest field, the sensor nodes act autonomously and their assignment depends on the application requirements. This application-dependent characteristic makes that a proposed solution (a routing protocol for instance) cannot be applied in all kinds of applications. In addition, WSNs are different from traditional networks and ad hoc networks in various aspects. The sensor nodes are resource-constrained devices. They have a limited storage and computation capabilities and are battery-powered. In the last two decades, proposing adapted solution to WSNs has attracted industrial actors and researchers. They address issues related to energy efficiency and data reliability in the data collecting process.
As a technique of artificial intelligence, machine learning (ML) is used to improve the network performance (Alsheikh et al., 2014)(Ma & Er, Meng, 2007)(Kim et al., 2021). Machine learning algorithms could be classified into three classes: supervised techniques, unsupervised techniques and reinforcement learning techniques(Alsheikh et al., 2014). The supervised ML algorithms require a labeled training data set is used to build the system model representing the learned relation between the input, output and system parameters. The goal is that the system learns the general rule that maps inputs and outputs. Unlike the previous category, the unsupervised ML algorithms determine, without labeled data set, the structure of the input. The goal is to classify the sample sets to different groups (i.e., clusters) by investigating the similarity between the input samples. In the reinforcement learning algorithms, the system learns by interacting with its environment (i.e., online learning).
Designing WSNs require to take in consideration resource constraints in terms of energy, computation and storage; of sensor nodes, dynamic topology and communication link failures. ML techniques are used to address divers challenges of WSNs functionalities such as energy-efficiency in routing, energy-efficient cluster formations (Forster & L., 2011) (Praveen Kumar et al., 2019).
Some of the applications of ML in WSNs are:
- •
ML for routing allows WSNs to learn from previous path establishment and to predict optimal routing actions and to perform dynamic routing to adapt the dynamic topology caused by nodes’ mobility, nodes’ energy depletion,..etc.
- •
ML for cluster formations and data aggregation allows WSNs to efficiently elect the cluster-head which significantly reduces energy consumption and enhances the network’s lifetime.
- •
ML for sink mobility in WSNs allows the mobile sinks to predict their trajectories to decrease the packet loss and enhance the network’s lifetime.
- •
ML for localization and objects targeting allows accuracy improvement.
- •
ML for Medium Access Control (MAC) layer. A MAC protocol controls the sensor node hardware responsible for interaction with wireless transmission medium (Bachir et al., 2010)(B. Yang et al., 2019). Reducing packet loss and prolonging the network lifetime can be achieved by developing resource efficient medium access control (MAC) protocols (Pasandi & Nadeem, 2019).