3D DAS VSP data were acquired in two active wells equipped with Electrical Submersible Pumps (ESPs) from the Big Foot field in the deep water, Gulf of Mexico. During the VSP survey, the wells remained producing most of the time with downhole ESPs operating at normal frequencies. Both seismic signals and ESP related noise were recorded by fiber optic cables installed inside the wellbore. In this paper, we compare DAS VSP data recorded with and without the ESP running and perform an analysis of the characteristics of flow noise activated by the ESP. We also use a machine learning (ML) algorithm based on supervised deep learning for attenuating ESP noise and present promising results.