Asymmetric flow velocity profile measurement using multipath ultrasonic meter with neural network technique

Average flow velocity in a multipath transit time Ultrasonic Meter (USM) is calculated by integrating the mean flow velocities along all of the acoustic paths. Every path has a weight that represents its contribution to the average velocity. In the conventional weighting methods, the flow for the entire USM is assumed to be fully developed and symmetrical, so that each path has a fixed weight of USM's measuring range. However, if the flow is not in the ideal conditions or distorted, implementing such conventional techniques will produce measurement results with poor accuracy. To handle that problem, an advanced weighting method is thus needed, so that a set of weights of all paths is able to adapt to several circumstances. In this paper, the performance of a conventional weighting method represented by the Covered Area (CA) is compared toan advanced weighting method represented by the Artificial Neural Network (ANN) using Root Mean Square Error (RMSE) analysis. In order to find the best architecture, several ANN parameters including learning rate, number of hidden layer, and neuron, are varied. The position and number of acoustic path are also varied based on the different number of USM Tomographic transducers from 10 to 16. The numerical simulation results show that the smallest RMSE for ANN with the Cascade algorithm is 0.0065 for USM Tomo 16 trans 6 paths, while in the case of CA is 2.2 10⁻³ for USM Tomo 12 trans 4 path. Furthermore, the sensitivity level of ANN-Cascade to the changes in number and path position is much lower than that of CA, whereas the range of ANN and CA RMSE are (6.5 10⁻³∼1.5 10⁻²) and (2.2 10⁻³ ∼7.6 10⁻²), respectively. By using tolerance within± 1% (AGA-9), the average RMSE for the testing sets is 9.1 10⁻³, while the ANN-Cascade and CA are 1 10⁻² and 1.8 10⁻².