A data-driven pipeline pressure procedure for remote monitoring of centrifugal pumps

Pumping systems are a key component of oil and gas pipeline transportation assets: monitoring their integrity is a crucial operation from a safety and revenue point of view. The solutions currently employed in the industry apply supervised machine learning techniques to data collected by multi-domain sensors directly installed on several positions of the pump itself; however, such approaches are not applicable on older machines, in contexts where a direct access to the pump is not possible, or whenever labelled data are not at disposal. This paper, instead, presents a predictive maintenance strategy where the condition of a centrifugal pump is tracked by solely exploiting standard pressure measurements, recorded also on remote points along the pipeline, and using an unsupervised learning approach. The smart monitoring strategy is presented and validated on historical pressure signals collected by Eni for several years on a crude oil transportation pipeline, located in Italy. Pressure data, recorded along the fluid line, are used to compute several statistical indicators on appropriate window lengths. These indicators are then fed to an unsupervised clustering procedure, based on a Gaussian mixture model. The output is an index within four different pump operational regimes, and a clustering visualization that permits the interpretation of the automatic regime classification. In fact, the manual inspection of the clusters shows that three of them describe standard modes (regular pumping operation, pumps off, flow regulations). The fourth one corresponds to high amplitude peaks in the signals and indicators, and so it is tagged as “anomalous” mode: pump maintenance logs reveal that the peaks are associated to damaged roller bearing movements, which disappear after the activation of the pump backup system. Anomalies are reported several days before the pump switch, so that a preventive maintenance could have been triggered. The robustness of the clustering algorithm is assessed on a statistical basis, whereas the overall validity of the monitoring system is tested on an instantaneous basis by applying the proposed model on two independent datasets, collected on real transportation pipelines: the results demonstrate the reliability of the proposed monitoring strategy in predicting and detecting all the pump failure events reported by the available maintenance logs. With respect to the mostly employed approaches, our machine learning procedure does not require any previous supervision of the data. Moreover, input data are the pressure transients produced by the pumps and guided within the fluid in the pipeline for long distances: pump failure analysis can be run using sensors located at several kilometers of distance from the pump itself, making a remote control strategy feasible.