Water quality prediction on a Sigfox-compliant IoT device: The road ahead of WaterS

Abstract

Water pollution is a critical issue that affects the entire ecosystem, with non-negligible consequences on humans’ health, thus inducing economic and social concerns. This paper focuses on an Internet of Things water quality prediction system, namely WaterS, that remotely communicates gathered measurements leveraging the Sigfox communication technology. The solution addresses the water pollution problem while considering the peculiar Internet of Things constraints such as energy efficiency and autonomy. The proposal demonstrateshow it is possible to detect and predict water quality parameters such as pH, conductivity, oxygen, and temperature by using WaterS, the Tiziano Project dataset, and a Deep Learning algorithm based on a Long Short-Term Memory recurrent neural network. The discussed water quality measurements are referred to the dataset that belongs to the Tiziano Project, in a period of time spanning from 2008 to 2012. The Long Short-Term Memory applied to predict the water quality parameters achieves high accuracy and a low Mean Absolute Error of $0.20$, a Mean Square Error of $0.092$, and finally a Cosine Proximity of $0.94$. The obtained results are analyzed in terms of protocol suitability of the current architecture toward large-scale deployments. From a networking perspective, with an increasing number of Sigfox-enabled end-devices, the Packet Error Rate increases as well up to 4% with the largest envisioned deployment. Finally, the source code of WaterS ecosystem has been released as open-source, to encourage and promote research activities from both Industry and Academia.

Introduction

The Internet of Things (IoT) is a well-known paradigm that turns devices into interconnected smarter objects. IoT devices are generally characterized by low computational power, networking limitations, and communication capabilities. As a matter of fact, IoT devices have to deal with issues related to data exchange while optimizing the communication protocols in terms of latencies, bandwidth, security, and energy consumption [1], [2], [3]. Despite their intrinsic limitations, IoT devices are now part of continuous monitoring processes in several fields, from industrial applications [4], to monitoring activities connected to air quality and environmental parameters in the most modern smart cities [5].

Among the many fields that could benefit from the introduction of IoT technologies, water quality monitoring is undoubtedly one of the most relevant and recently investigated [6], [7], [8], [9], [10], [11], [12], [13], [14]. In this context, WaterS [15] has been already proven to be able to provide remote monitoring capabilities for some of the most representative water quality indicators (i.e., temperature and turbidity) [16]. At the same time, the WaterS architecture provides an energy-harvesting and an ultra-low-power Sigfox-compliant radio interface to keep track of the continuous monitoring activities carried out by the IoT architecture. Even though the monitoring activities are of utmost importance to grant fine-grained sampling of the parameters of interest, water pollution, and other long-lasting phenomena could still be detected. In fact, since a large portion of the world’s freshwater lies underground, infiltration into the ground could be underestimated. Therefore, a system like WaterS could be more and more useful if it was able to carry out a prediction analysis of water quality parameters. Such an advancement could lead to the massive adoption of smart and energy-efficient sensing units to be employed in hostile areas, for example, subject to massive and pervasive pollution phenomena such as the spillage of toxic waste into the aquifers where water infiltration is a crucial task [17].

WaterS is proposed as a stand-alone, energy-efficient and standard-compliant system for measuring and forecasting water quality parameters.

The system was tested close to the seaside in the city town of Bari, Italy. The dataset involved in this study is one of the main outcomes of the Tiziano project [18], and it has been fully exploited to train a deep learning model for forecasting, i.e., in our case a Long Short-Term Memory (Long Short-Term Memory (LSTM)) neural network. WaterS has been developed by adopting open-source hardware/software (with a focus on the energy harvesting [19] capabilities of our solution) and a standardized wireless protocol to push the innovation further, as well as to allow researchers and academia to further use our code as a ready-to-use basis for further software development [20].

Contribution. This work aims at integrating an advanced deep learning technique, namely LSTM within WaterS to improve the current solution by providing additional features like the water quality prediction. Specifically, we provide an experimental evaluation where we show how the adoption of LSTM can effectively predict the reference data by assessing the results accordingly the Mean Square Error, Mean Absolute Error and Cosine Proximity metrics. In particular, the neural network has been configured to search for correlations on multivariate time series on surface water. Indeed, the experimental results demonstrated that some degree of correlation exists and this proves that it is worth pursuing estimations on the proposed variables. In addition, the achieved results show that the adoption of Recurrent Neural Network (RNN) for the analysis of water quality is a winning solution for the study of multivariate time series [21]. Comparisons against competing solutions show the viability and efficiency of our proposal. Finally, WaterS has been fully implemented as the first open hardware/software solution, and the source code has been released as open-source [20]. This permits the research community and companies to reproduce our results, use the solution on top of existing Sigfox transceivers, adopt the released code as a ready-to-use basis for further improvements and comparison and, finally, allow the interested readers to verify our claims.

Roadmap. The remainder of the present work is as follows: Section 2 is three-folded, since it introduces the reference background on (i) water quality monitoring IoT systems, (ii) Low-Power Wide Area Networks (Low-Power Wide Area Networks (LPWANs)) technologies, with a focus on the Sigfox protocol, and (iii) a thorough analysis on LSTM. Section 3 describes the operating scenario in which WaterS is adopted, as well as the envisioned architecture and the proposed prediction system. Section 4 summarizes both the leading criteria and methodological approach. On top of that, Section 5 presents the experimental campaign while Section 6 discusses the obtained results. Possible strategies for improving the WaterS systems are proposed in Section 7 together with the main findings, limitations, and future research directions. Finally, Section 8 tightens the conclusions and discusses future work directions.