Failure and anomaly detection for automatic meter reading network in intelligent water system

Abstract The article presents a solution for detecting failures and anomalies of the automatic meter-reading network in an intelligent water supply system. The main goal of the proposed system was to improve the maintenance of the water supply company. In the first stage, water movement from three locations in multifamily buildings and water pressure from the hydrophore station were collected, and then the collected data were aggregated and prepared for initial processing. In the next step, it was checked whether there were any gaps in the time series and, if so, they were supplemented by interpolation. Outliers were then detected and removed, using a local outlier factor. The time series prepared in this way were entered as input data to deep learning models. These models were used to obtain prediction values and prediction intervals for water flow and pressure values. Finally, the prediction error and a dynamic threshold of the prediction error value were calculated to signal a possible anomaly or failure. The obtained experimental results confirmed the effectiveness of the presented method.