Introduction
One fundamental challenge many facilities encounter is accurately estimating the rate of degradation throughout their facility. Degradation rates are used to inform facility technical leadership about the present risk of individual assets and are used to schedule a variety of inspection and maintenance tasks. Incorrectly estimating degradation rates can lead to an inadequate understanding of risk. If estimated rates are overconservative, facilities may waste resources on unnecessary inspections. Alternatively, if estimated rates do not correctly capture all potential risk, facilities can experience large economic or health, safety, and environmental (HSE) consequences.
Currently, degradation estimates are made by subject matter experts (SMEs) who leverage industry standard tools such as API RP 581, various industry-recognized damage/corrosion models, prior inspection data, and the wealth of knowledge and experience they have gained throughout their careers. While these methodologies have historically been a credible method of predicting degradation rates, actual rates can differ significantly from the estimated rates for a variety of reasons. First, while there may be a tremendous amount of data available to an SME, sifting through and analyzing a large quantity of data can be daunting for a human. Further, data quality issues such as incomplete, missing, or poor-quality data can also drastically alter SME perceptions. Finally, even with a complete and clean set of data, actual degradation rates can differ significantly from theoretical values due to a variety of factors such as environmental considerations and changes in facility process conditions.
Machine learning can strengthen the natural limitations of human SMEs, resulting in methods that predict degradation rates more quickly and accurately. When used properly, machine learning models can quickly sort through, organize, and clean massive amounts of input data such as temperature, pressure, metallurgy, and stream information, and leverage this data to make more accurate degradation rate predictions. Further, models based on data science can continually evolve and learn based on newly acquired data, preventing results from becoming stagnant.
Machine learning models can be leveraged – even by facilities with limited data – to strengthen areas of natural human limitations when predicting degradation rates. The following study found that a machine learning model was able to predict degradation rates for a hydrocracker unit more accurately and with a smaller margin of error compared to current industry practice. This article will discuss the details of this study as well as future applications of machine learning models for the industry.
The Hydrocracker Study
The purpose of this study was to compare the accuracy of degradation rates estimated by a machine learning model to the degradation rates calculated by human SMEs leveraging API RP 581. This study specifically focused on degradation rates for a variety of piping circuits in a hydrocracker. The study focused on piping circuits specifically because they are responsible for the majority of loss of primary containment failures in refining facilities [1].
The hydrocracker analyzed in this study had 25 piping systems which included 70 piping circuits and 1,662 condition monitoring locations (CMLs). The available data consisted of:
Comments and Discussion
There are no comments yet.
Add a Comment
Please log in or register to participate in comments and discussions.