Intelligent Ultrasonic Flow Measurement Using Linear Array Transducer With Recurrent Neural Networks

Abstract

To realize high-quality transit-time ultrasonic flow measurements, accurate and precise estimates of the transit-time difference are essential. In this study, we propose deep learning-based neural network (NN) models to measure the transit-time difference in an ultrasonic flowmeter using a linear array transducer. Three approaches to compute the transit-time difference are presented: the cross-correlation with phase zero-crossing (XCorr), fully connected NN, and recurrent neural network (RNN) with long short-term memory (LSTM). The training data for the NN were generated by simulating target time differences by utilizing the experimental data acquired in the pipe system. To evaluate the performance of the proposed methods, linear regression, the Bland-Altman plot, and the root mean squared error (RMSE) were analyzed using testing data from the experiment. The results of this study show that the RNN-based approach yielded improved performance with an accuracy of up to 94% and a 33.48% reduction in the RMSE, compared to the XCorr method. In addition to the time difference estimation, our proposed RNN-based model can replace the entire flow rate estimation process, including interpolation, velocity correction, and zero-flow calibration. This study demonstrates the feasibility of an intelligent ultrasonic flowmeter employing the RNN-based model with potential in industrial applications.