Modeling of degradation of pipes in refineries using machine learning

Important results from the project

We did not reach the objective of predicting pipe degradation based on crude oil feed, mainly because most of the available data could not show sufficient wear for modeling purposes. The project instead became a collection of experiences that may be of importance for continued data collection. In the preparatory aim for data from feeding renewable raw materials, this work can be seen as a necessary first step.

Expected long term effects

We did not succeed in developing models that, based on given crude oil feed, can predict pipe degradation with certainty. In general, this is due to the lack of wear and tear where the sensors are placed. However, there is an exception in a sensor where there is clear wear in a specific sensor, but probably another unanalyzed factor dominates the reason behind this. This means that the connection between feeding for specific crude oil types and the wear becomes uncertain, and thus the predictive power becomes low.

Approach and implementation

In the project, FCC has analyzed pipe thickness measurements from Preem obtained over the last 3 years. However, the overall lack of noticeable wear makes accurate modeling almost impossible, and in the only observed case of wear, it is not possible to link crude oil feed to wear with certainty. A desirable approach might instead be to have a more continuous iteration between collection and analysis, to improve data quality directly in the collection but with a starting point in the modelling.