Abstract
In this paper, three alternative solutions to the problem of detecting and cleaning anomalies in soil signal time series, involving the use of artificial neural networks deployed on in situ data measurement end devices, are proposed and investigated. These models are designed to perform calculations on MCUs, characterized by significantly limited computing capabilities and a limited supply of electrical power. Training of neural network models is carried out based on data from multiple sensors in the supporting computing cloud instance, while detection and removal of anomalies with a trained model takes place on the constrained end devices. With such a distribution of work, it is necessary to achieve a sound compromise between prediction accuracy and the computational complexity of the detection process. In this study, neural-primed heuristic (NPH), autoencoder-based (AEB), and U-Net-based (UNB) approaches were tested, which were found to vary regarding both prediction accuracy and computational complexity. Labeled data were used to train the models, transforming the detection task into an anomaly segmentation task. The obtained results reveal that the UNB approach presents certain advantages; however, it requires a significant volume of training data and has a relatively high time complexity which, in turn, translates into increased power consumption by the end device. For this reason, the other two approaches-NPH and AEB-may be worth considering as reasonable alternatives when developing in situ data cleaning solutions for IoT measurement systems.