Monitoring pumping units by convolutional neural networks for operating point estimations

Abstract

To avoid the failure of pumping units, the monitoring of operating points with a subsequent assessment of the condition of the pump may support the decision for required maintenance. For that purpose, convolutional neural networks (CNNs) are implemented to predict the operating points of pumping units. Instead of using traditional flowmeter and manometer, vibration and acoustic signals are used to estimate the head and volume flow rate. An appropriate pre-processing of raw data is applied, enabling our method to predict well on different datasets. For the datasets measured in an anechoic chamber, the best model of each subset achieves relative errors smaller than 4.9% for the prediction of head and 7.6% for the volume flow rate. For cases where only small amounts of data exist, it is furthermore demonstrated that transfer learning from one dataset to another dataset provides an improvement in performance.