In this study, we developed a mechanism-assisted data-driven model to regulate substrate feedback to improve the production efficiency of sophorolipids (SLs). First, we used a variety of on-line biosensors to establish a multi-scale parameter detection system. We found that the production of SLs by fed-batch fermentation could be divided into three stages: a stage that was limited by cell production capacity, a stage that was inhibited by high product concentration, and a stage that was limited by oxygen supply. Subsequently, we used process parameters to develop a data-driven model, and this was then combined with the analysis of cell metabolic mechanisms. The optimal production of SLs was achieved in the first and second stages by the precise feedback regulation of substrate feeding, which increased the titer of SLs by 4.9%. The control error of the substrate was reduced from more than 15% to less than 5%. The mechanism-assisted data-driven model was then applied for semi-continuous fermentation during the production of SLs. This effectively alleviated the oxygen limitation during the third stage, and further increased the productivity of SLs to 2.30 g/L/h, 40.2% higher than the fed-batch fermentation method.