Prediction of inlet gas volume fraction of rotary vane pump under variable operational conditions with 1DCNN-TL model utilizing vibration signals

The Rotary vane pump (RVP) exhibits promising potential in the lubricating oil system of aircraft engines. The entrainment of air into the lubricating oil flowing through the pump affects the suction and discharge flow rates of the RVP. Therefore, understanding the inlet gas volume fraction (IGVF) of the RVP is crucial to assess its capability to deliver sufficient lubricating oil for ensuring the safe operation of aircraft engines. This study presents a one-dimensional convolutional neural networks (1DCNN)-based transfer learning (TL) (1DCNN-TL) prediction model to predict IGVF for the RVP using experimental data from gas-liquid two-phase flow in the RVP. The model is pre-trained using the vibration signals under a specific differential pressure between the pump inlet and outlet to enable predictions of the IGVF under varying differential pressures. The 1DCNN serves as pretrained model for analyzing the outcomes of different transfer tasks. The effect of different frozen layer numbers on training time and prediction accuracy of the neural network is explored. Subsequently, the influence of varied source domain data on prediction accuracy is investigated, revealing that for optimal TL results, stable condition data at 200 kPa should be preferred as the source domain. Then the 1DCNN-TL prediction model of IGVF of RVP under variable operational conditions is determined. The results indicate varying iteration convergence rates of the pre-trained model under a differential pressure of 200 kPa across different transfer tasks, with a final goodness of fit exceeding 0.958 after fine-tuning the model. The model demonstrates the high prediction accuracy and applicability, which can contribute to improving the reliability and safety of aircraft engines.