Flow Soft Sensing of Hydraulic Pump With Model Updating Considering Wear Degradation

Flow soft sensing of hydraulic pump is a crucial way to avoid use of flow sensor, but it is difficult to achieve due to the high nonlinearity relationship with pressure, temperature, and rotational speed. Especially, the nonlinear flow relationship changes gradually with wear degradation. To cope with this issue, a flow soft sensing method is proposed combined with transfer learning and semi-empirical modelling in this study. A soft flow sensing model was firstly trained using sufficient dataset based on Back Propagation Neural Network (BPNN). An iterative model updating method is proposed with parameter-based transfer learning to cope with wear degradation after long-term operation, in which semi-empirical mathematical modelling is involved to provide an additional constraint and reduce the requirement of large sample size. In the updating process, the weights and thresholds of the initial soft-sensing model are fine-tuned to reduce the flow prediction error in forward propagation. The experimental results show that compared with the initial BPNN model and the updated model with transfer learning directly, the proposed method can reduce the average flow sensing error from 1.43 L/min and 3.25 L/min to 0.34 L/min, and the standard deviation (SD) from 0.57 L/min and 3.31 L/min to 0.35 L/min. Thus, the proposed method can achieve accurate soft sensing of the pump flow under wear degradation effectively.