Flow sensors substitution using neural networks with small data for external gear pumps

ABSTRACT This study introduces a data-driven model that enables the direct replacement of volume flow sensors in external gear pump applications. As this study aims to meet the requirements of real-world industrial scenarios, it is necessary to utilize a small data approach to demonstrate an applicable and scalable solution. The method improves through a data augmentation process based on fundamental physical laws, reducing the need for an extensive data set. A neural network predicts the volume flow within the pump's operating points over its entire operating range. The control architecture, including failure mechanisms, is presented, and the execution time is validated under real conditions on the microcontrollers. Two different fluid systems are investigated with three different types of external gear pumps and validated over the entire operating range of the fluid machinery. The methodology achieves a mean absolute percentage error of 1.53% considering the output volume flow of the pump systems.