Advanced Analytics to Maximize Operations Continuity and Optimize the Asset Value: The Fouling Real Example
- G. Camarda (Eni SpA Upstream and Technical Services) | L. Cadei (Eni SpA Upstream and Technical Services) | M. Montini (Eni SpA Upstream and Technical Services) | G. Rossi (Eni SpA Upstream and Technical Services) | P. Fier (Eni SpA Upstream and Technical Services) | A. Bianco (Eni SpA Upstream and Technical Services) | A. Corneo (Eni SpA Digital Competence Center) | D. Milana (Eni SpA Digital Competence Center) | D. Loffreno (Eni SpA Digital Competence Center) | L. Lancia (Eni SpA Digital Competence Center) | M. Carrettoni (Eni SpA Digital Competence Center) | G. Silvestri (Eni SpA Digital Competence Center) | F. Carducci (The Boston Consulting Group GAMMA) | G. Sophia (The Boston Consulting Group GAMMA)
- Document ID
- Offshore Mediterranean Conference
- Offshore Mediterranean Conference and Exhibition, 27-29 March, Ravenna, Italy
- Publication Date
- Document Type
- Conference Paper
- 2019. Offshore Mediterranean Conference
- 6 in the last 30 days
- 35 since 2007
- Show more detail
Fouling events in critical machinery as re-boilers often lead to sudden production blocks. In-advance identification of the first symptoms can provide plant operators with enough time to prepare for the upcoming events, avoiding halts and therefore optimizing the overall production.
Re-boilers in particular follow a distinctive life cycle: they are usually characterized by a steady phase, as the fouling is a long process whose effects can only be noticed after a relevant number of years. Hence, even for a very broad period there is only a limited amount of relevant events, which is in turn a widely-recognized constraint for any data-driven approach. Moreover, when the fouling becomes noticeable, the situation can deteriorate very quickly. Thus, identifying in advance the first symptoms is of paramount importance, as it allows to gain more time to intervene without penalizing the ordinary site operations, thus guaranteeing operations continuity.
In order to achieve this objective, we follow a novelty detection approach: by training our model on the data that belongs to the standard operating mode of the equipment (i.e., learning its normal behaviour), we expect it to learn to detect anomalies in the future. First, we define a set of synthetic features that summarize enough information from the signals. Then, using a One-Class Support Vector Machine, we are able to distinguish normal from abnormal patterns. This means that the model must be frequently retrained, to be able to adapt easily to the different working conditions (such as a change in the layout of components in the field or the introduction of different chemical agents). Therefore, we follow a training-predicting iterative approach, where we take advantage of the user feedback to adapt to new operating conditions, such as an evolution in the organization of wells or lines or a change in scale.
Results show that even in presence of a physical process that produces strong limitations in the data available and in the number of relevant events, it is possible to leverage unsupervised machine learning techniques to automatically diagnose anomalies and risks, allowing to protect key productive assets.
The Anomaly Detection System developed for re-boilers in gas treatment production lines is a digital monitoring system able to identify early signs of fouling. The project has been developed applying the agile approach, integrating skills and know-how of different professionals, from data science to process and production engineering, operations and ICT.
|File Size||798 KB||Number of Pages||9|