Deep learning models for data quality classification in water systems.

Water systems around the world are under increasing pressure year by year. Large numbers of the world's population currently live in water-stressed areas, and with current trends that will only get worse. Using smart city technology is a way to optimise current systems, using water meters to continuously record information about consumption levels and measurement instruments within the pipes to convey data. In this thesis we will treat the problem of cleansing the data coming directly from the pipes, as it is a hazardous environment with a large propensity for damage to the instruments. The main issues are sensor fouling, leading to incorrect velocity measurements, and pressure compensator malfunction, leading to incorrect level measurements. Other problems can be from decreased performance with increasing age, or incorrect values around the limits of measurement. The goal is to predict labels for the data samples based on the velocity measurement, so that the low quality data is excluded from the datasets and only the high quality data is used in the smart city process of creating a digital twin of the system. The data used is from three different measurement campaigns on systems in the north of Italy, spanning a period of three years. We put a large focus on data preparation, scaling, and subdivision of the datasets for exploiting the maximum information available, taking into careful consideration the time element of the data. The measurement points from the three different campaigns are divided such that we can have three testing sets for the model, each providing insight into the learning. We train and optimise a convolutional neural network, achieving final accuracies of over 90% on the various test sets.