Condition Monitoring Approach Based on Unsupervised Anomaly Detection for Pumps Regulating Groundwater Level Under a Coastal Infrastructure

This paper describes a condition monitoring approach for pumps regulating groundwater level under a port infrastructure. We focus on the Bremerhaven container terminal located in northwest Germany at the mouth of the river Weser. Our aim was to construct a strategy to detect potential pump failure indications that could inform conditional maintenance actions. Two signals were available for us: the groundwater level, measured with a radar, and the binary pump on/off operation signal. For this purpose, we tested four unsupervised machine learning-based anomaly detection algorithms, in combination with multiple post-processing methods for anomaly scoring and thresholding. Additionally, we developed a model to simulate the groundwater level signal, enabling the test of failure modes that were not present in measured data. We found that the appropriate selection of model and post-processing method was critical for obtaining satisfactory results in both measured and simulated signals.

Keywords: Condition based monitoring, Anomaly detection, Deep learning.