Elsevier

Computer Communications

Volume 193, 1 September 2022, Pages 155-167
Computer Communications

Cognitive quality of service predictions in multi-node wireless sensor networks

https://doi.org/10.1016/j.comcom.2022.06.042Get rights and content

Abstract

Wireless sensor networks lead the way to the realization and penetration of the internet of things into daily life. As the sensors and communication devices interconnect things and people more, the quality of service demands become stringent and diverse. Optimizing conflicting quality of service goals is NP-hard. Moreover, the need for the communication systems to dynamically adapt has grown as factors like scale, application-specific performance demands, and deployment scenarios evolve. In this article, we intend to predict the quality of service with an aim to improve the performance of services in the internet of things. We conducted experiments using a real test-bed and collected performance data under a wide range of communication parameter configurations. Statistical analysis revealed a significant relationship between communication parameters and quality of service metrics. Based on the correlations, we trained deep learning models to assess the predictability of the metrics. The prediction results are encouraging. The outcomes of this research pave the way to lay the foundation for a data-driven design of adaptive quality of service in wireless sensor networks and the internet of things.

Introduction

The ubiquitous presence of Wireless Sensor Networks (WSNs) has contributed greatly to the realization of the internet of things (IoT) [1], and this increasing scale also positions IoT in more challenging situations. WSNs find applications in many areas of daily life where sensor nodes are used to carry out specific tasks without the intervention of humans. Some of the application domains include agriculture, medical units, underwater monitoring systems, smart home appliances, smart industries, etc. [2], [3]. Lately, wearable sensor technologies have caused a surge in the number of people connected to the internet. [4]

The evolution in WSNs has diversified its application domains and hence meeting performance goals become challenging. It is common to find particular Medium Access Control (MAC) and physical (PHY) layer solutions to meet the performance demands of specific application domains. However, devising MAC and PHY protocols leads to heavy experimentation, standardization, and manufacturing costs. The costs come in the form of monitory as well as time. The other major limitation of the aforementioned approach is the lack of adaptivity. The protocols are usually hard-defined with minimal configurability and cannot self-tune as the network and application scenarios exhibit evolution. We can address the problem of adaptivity by maximizing the configurability of various parameters protocols by identifying their respective relationships with the quality of service (QoS) metrics as well as selecting appropriate MAC protocol. It is pertinent to mention that configurability of parameters means assigning appropriate values to different parameters at different layers of the protocol stack in use (e.g., packet size, traffic rate, transmission power, backoff, etc.). Some of these parameters (which are protocol specific) can contribute to the configuration of a protocol. Whereas, we also propose to choose a particular protocol depending on the QoS need. For example, the current implementation of Contiki operating systems allows for choosing CSMA or TSCH as a MAC layer protocol. Both have different characteristics and can behave differently under different conditions. Understanding the relationships among the communication parameters and QoS metrics is vital to meeting the conflicting QoS goals in changing network conditions and emerging application requirements. [5].

Research in the domain of WSNs has been extended to diverse considerations. Work presented in [6] highlights the importance of efficient wireless sensor and actuator node organization in critical sectors like agriculture. The proposed solution is not based on any adaptive method like machine learning. Considering the growing importance of the cloud and integration of WSNs with it, [7] proposed a solution to reduce the network load being transferred to the cloud by facilitating processing at the edge. The proposed solution can benefit from possible adaptive network-level configurations. A solution based on vertical handover operations for tackling the QoS problem in wireless networks is presented in [8]. A machine learning-based endeavor is presented in [9] that discusses efficient QoS in industrial wireless sensor and actuator networks using support vector machine. The proposed solution focuses on interference classification that can be helpful in further characterizing QoS.

Machine Learning (ML) models can play a vital role in predicting the optimal configurations for communication parameters that can meet the required QoS goals. ML models are trained under different stack parameter combinations to predict the performance metrics. We propose a solution for Packet Delivery Ratio (PDR), Throughput (THP), and energy consumption (radio and CPU) by considering Transmission Power (TP), Maximum Transmissions (MT), Distance (DT), Packet Size (PS), Inter Arrival Time (IAT) and Number of Nodes (NN). It is pertinent to mention that MT represents the maximum transmission attempts allowed until a packet is successfully delivered, whereas IAT refers to the gap in milliseconds between successive packet transmissions. To train the dataset, we used Deep Neural Network (DNN). This multi-parametric approach has been found in literature [5], [10]. The purpose is to understand the relationship of these parameters to QoS metrics. Results achieved from this experiment give sufficient insight into the dynamics of communication parameters and their influence on QoS performance metrics.

ML models are widely used to create prediction models for many WSNs application scenarios. A link quality prediction technique is presented in [11], based on the naïve Bayes classifier, logistic regression, and artificial Neural Networks (NN). The proposed model uses Packets Received Percentage (PRP) and PHY layer information as input parameters and predicts the reception of the next packet as output. Logistic regression performs better with a small computational cost. The main limitation of this work is that too few parameters are used to capture any meaningful variations and relationships among QoS metrics and parameters of interest. In [12], a data-driven technique is presented for MAC layer optimization by considering real-time communication information. Again, it follows the common limitation of using limited parameters as well as adopted receiver side quantities to predict sender side parameters which are impractical. Authors in [13] used offline mode and traces for performance evaluation. Mathematical programming techniques are highly complex for constrained devices like sensor nodes. Neural networks based prediction models for predicting network lifetime, TP, and DT between nodes are proposed. The proposed approach is purely a simulation and shares the lacking discussed for [12]. In [14], an attempt for collaborative QoS prediction is presented. This technique is based on a context-sensitive matrix-factorization approach for prediction. It considered both implicit and explicit factors in QoS data, thus, providing better insight for QoS prediction.

There are two major limitations of the above-mentioned solutions. First, most of the solutions use very few parameters that lack capturing any meaningful relationship among QoS metrics and parameters. Second, many proposed schemes rely on using receiver-side parameters to predict sender-side quantities. In addition, some solutions are based on simulated data which makes it hard to achieve representative results for real-world situations.

Authors in [15] also provide ML-based QoS solutions for predicting QoS parameters for better user-side experience in both WSN and IoT applications. In [16], authors proposed a neural network model for QoS prediction. In [17], a goal-driven service model for the QoS prediction model is presented for IoT which is adaptive and self-organizing. An overall limitation in the literature summarized above is that the number of parameters used to train the ML models is limited. Another shortcoming that these studies suffer from is that in many cases receiver-side parameters are used as input features at the sender-side for prediction. The common limitation in these solutions is that the QoS is predicted at a service level and no attention is paid to the network-specific settings that keep the potential of accurate and adaptive predictive system abstract.

In recent endeavors [18], [19], [20], [21], we extended the number of parameters for predicting QoS metrics as well as avoided the use of receiver-side parameters as features for sender-side predictions. In [18], PDR and energy consumption are predicted using more than 48 thousand combinations of as many as 7 pre-configured parameters (including IAT, PS, TP, DT, MT, etc.). In [19], the delay is predicted using the same parameters. PDR with signal-to-noise ratio is predicted in [20]. The major limitation of results presented in [18], [19], [20], [21] is that the experimental data were collected using a single-hop topology. The current work focuses on the experimental data collected under multi-node scenarios.

The rest of this paper is organized into 4 sections. The experiment details including topology and parameter configurations as well as the adopted ML method are described in Section 2. Section 3 elaborated on the statistical analysis between communication parameters and QoS metrics. Prediction results are detailed in Section 4. Section 5 concludes the paper and highlights the significance of the research work as well as future leads.

The limitations that motivate the present research can be summarized as follows:

  • Meeting conflicting QoS goals in prevailing deployments of WSNs and IoT requires solving complex optimization problems which are often NP-hard.

  • Conventional methods to improve QoS lack in adapting to evolving network parameters.

  • Configurable parameters provide an opportunity to fine-tune those according to the current need of the QoS.

Based on the limitations and motivations the current research is intended to contribute the following:

  • Analyze the metrics like packet delivery ratio, throughput, and energy consumption as various communication 56parameter settings impact these metrics.

  • Use machine learning models to identify how each feature and different subsets of features contribute to determining the accuracy in predicting all the metrics considered.

  • Rank the features that yield the best prediction results in the context.

The proposed work potentially lays down the foundation to realize an adaptive framework to facilitate QoS in the evolving IoT settings.

Section snippets

Materials and methods

The idea of this experiment was to carry out foundation work to build an intelligent framework for QoS predictions in WSNs. QoS Predictions in WSNs have gained attention recently.

In our previous efforts [18], [19], [20], we have predicted metrics like packet delivery ratio, energy consumption, signal-to-noise ratio, and delay using various communication parameters as shown in Fig. 1. We adopted various machine learning models and used various communication parameter settings to train and test

Statistical analysis

In the following, an pictorial analysis of the relationship among the communication parameters and QoS metrics is presented.

Correlations among parameters and metrics

Statistical correlations are described in Fig. 8. It shows the R2 scores of the relationship among the parameters/features (y-axis) and QoS metrics (x-axis). The magnitude of correlation scores of MP, TP, IAT, NN, DT, and PS with THP are 0.77, 0.77, 0.59, 0.53, 0.35, and 0.35, respectively. In the case of PDR, DT (0.55) proves to be most correlated, followed by TP (0.27), and NN (0.26). Other metrics that tend to have strong correlations include time in transmission (TiTx), time in DLPM

Conclusions

Modern communication systems are envisioned to be adaptive. The advent of software-defined networking and network function virtualization makes it possible to implement adaptive designs. Such adaptive design can be realized using dynamic fine-tuning of configurable communication parameters with the help of machine learning. In this paper, we demonstrated the potential of using machine learning for predicting quality of service metrics in wireless sensor networks and the internet of things based

Funding

This research was supported by the Higher Education Commission of Pakistan under National Research Program for Universities (NRPU) 7522/Balochistan/NRPU/R&D/HEC/2017.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgment

This work was supported by the National Research Foundation of Korea , under project BK21 FOUR.

References (22)

  • GrimaldiS. et al.

    An SVM-based method for classification of external interference in industrial wireless sensor and actuator networks

    J. Sens. Actuator Netw.

    (2017)
  • View full text