Digital Transformation of Historical Data for Advanced Predictive Maintenance

Abstract

In recent years, big data technologies have paved the way for digital transformation in oil and gas industry. Multi-domain measurements are collected by advanced sensor systems and processed using data-driven approaches, allowing to derive constitutive relations between the operational status of the asset and the measured variables. In addition, historical pressure measurements can be exploited for advanced pipeline monitoring. This paper presents a methodology, applied to a case history, where legacy data are repurposed and employed both to track pump health and to enhance the digital conversion. The dataset consists of past pressure signals collected by Eni for several years at the pumping terminal of a crude oil transportation pipeline, which has a length of 100 km and 16" diameter pipes, located in Italy. Pressure transients' variance, kurtosis and variation range, computed on appropriate window lengths, are fed to an unsupervised clustering procedure based on a Gaussian Mixture Model (GMM), which automatically identifies four clusters. An expert analysis of the labeled data reveals that each cluster corresponds to a well-defined and different pump operational mode, namely: standby (pumps off), transition (pumps switching on/off), normal (line flowing) and anomalous. The latter mode is connected to a high value in the pressure transients' variance and kurtosis: during such regime, pump maintenance logs report a failure and replacement of a system part. Interestingly, the anomalous condition starts to show up several days before the actual part replacement. The proposed case history reveals the potentiality of: adding value to legacy data, as they can be reprocessed, tagged and used as supervised examples in the training phase of new data-driven procedures; comparing, merging and complementing monitoring strategies of assets at different digitalization stages; aiding the development of predictive maintenance strategies.