Neural-network-based error reduction in calibrating utility ultrasonic flow meters

Abstract

In this paper, a Multi-Layer Perceptron Neural Network (MLPNN) based calibration is proposed for a utility ultrasonic flow meter. The measured flow range is from 0.2 to 4 m3 per hour. Due to the presence of transient flow in a significant portion of the flow range, as well as very low flow velocity in the laminar flow region, nonlinear effects appear that cause a variable bias error with fluid flow changes. Therefore, in order to reduce the bias error and increase the accuracy of the flow meter measurement, the use of neural network is considered in the flow meter calibration due to its nonlinear mapping ability. Neural network training is done based on input and target data. The sing-around method is used to measure signal time of flight. In this study, the goal is to achieve an error of less than 1.5%. To evaluate the suggested method, two-point calibration and bracketing method were also used and the results were compared. Since calibration equations is a mapping between the flow meter and prover or master meter reading, the objective is to reach a method that is able to predict or map the flow rate with a smaller error. Although the two-point calibration method resulted in good statistical agreement between the flow computer output and base values at data collection points, it failed in attaining the above objective, but the bracketing technique worked with a prediction error slightly better than 1.5%. On the other hand, using a MLPNN produced a maximum prediction error of 1.21%, and thus, the best output. This shows that the use of Artificial Intelligence as an efficient calibration tool is worth further exploration.