THE ROLE OF SURROGATE MODELS AND MACHINE LEARNING IN MODERN ULTRASONIC FLOW MEASUREMENT

Abstract

Recent developments in the field of machine learning have found their application in a wide range of design processes. They have particular use where numerical simulations are involved and fast, more accurate predictions and optimized models are very much needed. In order to speed up experiments on a device or system model, it is necessary to speed up its execution (simulation). Instead of detailed models, you can create a surrogate. Its main task is fast execution, small amount of occupied memory and preservation of the specified error threshold in relation to the detailed model. This article demonstrates the integration of machine learning into the flow measurement process using ultrasonic flowmeters. The main source of errors in the application of the modern ultrasonic principle of flow measurement arises from the difficulty of taking into account the actual velocity profile of the measuring flow. In practice, the distribution of velocities in the cross-section of the pipeline differs from the theoretical one introduced in the calculation algorithm. However, if the velocity profile is known, an appropriate correction can be estimated and taken into account during calibration. This will increase the accuracy of measurements. In this study, an intelligent compensation of errors caused by profile distortion was presented to improve the accuracy of using multipath meters in such ultrasonic conditions. The purpose of such an intelligent correction arises in the search for the optimal layout and the minimum sufficient number of chords in the measuring transducer for various installation conditions. The adoption of a new approach based on a surrogate model with a neural network made it possible to take an approximate flow profile that has a certain distortion. So, for the chosen topology of the acoustic flow sensing channels, programmatically, by changing the location angle of the measuring system, instead of the local resistance, add such a position of the chords for which it is possible to set maximum admissible measurement accuracy. This means using a neural network for the required input correction model, especially in an environment characterized by a change in the velocity profile under the influence of different flow distortions.