Machine learning identification of multiphase flow regimes in a long pipeline-riser system

Abstract

Multiphase flow regime identification is a promising technology for ensuring flow safety in marine gathering pipelines. An experimental study of air-water flow regime was conducted on a 1687 m long S-shaped pipeline-riser system. The flow regimes were quantitatively classified as severe slugging, oscillating flow and stable flow. By combining of two cheap and easily accessible differential pressure signals on the top of the riser as a sample, the flow regime classifier is established based on the support vector machine, and methods to improve the computational efficiency of the classifier are investigated. First, on the premise that the recognition rate is over 90%, a reasonable sample-size reduction strategy based on the K-means clustering method was designed. When the signal duration was 18.6 s, the number of samples was reduced from 7632 to 2658, and the hyperparameter iteration time of the classifier was shortened by 99.3% from 10681 s to 80 s. Second, with the combination of the single-feature recognition rate and the correlation between features, the number of features was reduced significantly. The recognition rate of the three preferred features was 96.3% with a sample duration of 18.6 s. Finally, by using the samples with a signal duration of 18.6 s as the training set, the flow regime classifier built with the three preferred features had a better generalization ability, and the average recognition rate of the testing set was higher than 90%. When the signal duration of the training set is in the range of 9.3 s–55.8 s, the maximum difference of the average recognition rate of the testing set is only 0.7%.