Multi-target Regression Approach for Predictive Maintenance in Oil Refineries using Deep Learning

Modern oil refineries typically use a high number of sensors that generate a massive amount of data about various process variables in the infrastructure. This data can be used to perform predictive maintenance, an approach to predict impending equipment failures and mitigate downtime in refineries. This paper presents the use of multi-target regression approach for predictive maintenance. Multi-target regression is a modeling approach that aims to predict multiple targets simultaneously. The relationships between multiple process variables are modeled using deep learning methods, while the model error is evaluated using cumulative sum method to detect faults that might potentially become failures. Unlike many existing solutions, our approach does not rely on the availability of data that captures the presence of faults in the plant. The proposed approach is demonstrated using real industrial data from a crude distiller in Shell Pernis. The results show a speed-up in modeling time by 16x and an improved early fault detection time by 1.2x as compared with the single-regression approach. Furthermore, the proposed approach is also able to isolate the faults by producing higher errors in predicting faulty equipment compared with healthy equipment.