Flow Velocity Measurement in Dredging Pipeline Based on Transfer Learning and Electrical Resistance Tomography

Flow velocity is a critical parameter in dredging and transportation pipelines. Velocity measurement methods using electrical resistance tomography (ERT) have attracted increasing attention. A typical ERT sensor has a caliber of 800 mm. At this caliber, a convolutional neural network (CNN) can accurately calculate flow velocity after being trained on sufficient historical data and expertise accumulated over time. To meet different operational requirements, we have deployed ERT sensors with various calibers, but we can utilize little existing historical data and expertise to measure flow velocity. To address this issue, we propose a transfer learning (TL) model to construct a CNN in which we optimize the loss function by incorporating the multikernel maximum mean discrepancy (MK‐MMD) between the source and target domains. Furthermore, through multi‐task learning, the proposed model mitigates negative transfer and inadequate feature learning. Experiments in dredging engineering demonstrate that the proposed model effectively leverages existing historical data and achieves high reliability and accuracy in flow velocity prediction across a set of evaluation metrics.