Artificial Neural Networks for Real-Time Data Quality Assurance

Abstract

Wireless Sensor Networks used in aquatic environments for continuous monitoring are typically subject to physical or environmental factors that create anomalies in collected data. A possible approach to identify and correct these anomalies, hence to improve the quality of data, is to use artificial neural networks, as done by the previously proposed ANNODE (Artificial Neural Network-based Outlier Detection) framework [1]. In this paper we propose ANNODE+, which extends the ANNODE framework by detecting missing data in addition to outliers. We also describe the design and implementation of ANNODE+, implemented in Python to exploit readily available machine learning (ML) tools and libraries, also allowing online processing of incoming measurements. To evaluate the ANNODE+ capabilities, we used a dataset from a sensor deployment in Seixal's bay, Portugal. This dataset includes measurements of water level, temperature and salinity. We observed that our implementation of ANNODE+ performed as intended, being able to detect injected anomalies and successfully correcting them.