Corrosion prediction in oil and gas pipelines: A machine learning approach

Oil and gas pipeline corrosion is highly dangerous for the environment and for the human being. In fact, this unlucky event can cause oil leakage in the environment through the rupture of the pipeline itself, while, depending on the environmental condition in which the pipeline operates, gas leakage can cause severe explosions. It is hence critical to design prediction models of the presence of corrosion in order to improve prevention and control [1]. In recent years, due to the steadily increasing availability of data, Machine Learning (ML) techniques are becoming a standard de facto in lots of different applications. Properly exploiting these techniques in the field of pipeline corrosion would pave the way to a huge improvement in the prevention and control of this dangerous problem. Unfortunately, pipeline infrastructures are not endowed with fluid-dynamical sensors, which means that an important family of descriptor for this dangerous phenomenon are not available without using simulations. This issue raises the problem of integrating and validating simulated information with real measurements of other descriptors (mainly geometrical) of the phenomenon. Therefore, in this paper, we will show how we can integrate these two different sources of data in order to build a rich and descriptive dataset representing the corrosion phenomenon in a pipeline infrastructure. We will then analyze, through a cross-correlation analysis, the relationships existing between the variables within our dataset to verify the quality of the integration procedure. Finally, we will seek for some insights in the pipeline corrosion phenomenon through a feature selection procedure, which will allow us to check whether some attributes of our dataset are privileged descriptors of the presence of corrosion in a pipeline.