Author(s): Rocco Palmitessa; Erling Amundsen; Jesper Mariegaard

Linked Author(s):

Keywords: Nomaly detection; Data imputation; Urban drainage; Sensor data; Machine learning; LSTM

Abstract: Information and communication technologies combined with in-situ sensors are increasingly being used to enable smart control of urban drainage systems. However, sensor data streams may become compromised due to sensor failures, system anomalies, or malicious tampering, jeopardizing the reliability of data-driven control systems. Automatic anomaly detection can identify these data quality issues early to improve system reliability, but the robustness of data-driven methods to future unknown anomalies is uncertain. In this study, we tested the performance of Long Short-Term Memory (LSTM) neural networks for anomaly detection and data reconstruction on a large open dataset of water depth from a combined drainage system. Synthetic anomalies of seven different types were injected into a test dataset to mimic realistic sensor faults of varying intensity and duration. The proposed data validation workflow correctly labeled 82% of the synthetically altered events, with frozen signals and low intensity anomalies being the most difficult to detect. When these are excluded, the accuracy increased to 96%. The workflow also reconstructed altered data to within 1.5 cm mean absolute error. These results demonstrate the robustness of automatic validation for real-world applications.

DOI: https://doi.org/10.64697/978-90-835589-7-4_41WC-P2010-cd

Year: 2025