Bayesian machine learning for virtual flow metering

Abstract

Machine learning is emerging as a promising approach broad range of modelling applications. One of these applications is soft-sensing in industrial processes. Soft-sensors are mathematical models that predict process-related quantities that are otherwise difficult or expensive to measure. Additionally, it can be difficult to derive mathematical models for these quantities from physics, which is where machine learning enters the picture. Machine learning allows us to obtain predictive models from large quantities of data. However, there are not always large quantities of informative data available, due to the way industrial processes are operated. We therefore seek to expand our data foundation, by learning from data collected from other processes, similar to the ones we are interested in. This allows us to model difficult phenomena with little and still achieve a high level of performance, due to the ability to leverage experience from other related processes. The primary process studied in this work is the multiphase flow through valves, but other applications have also been explored with promising results.