| The digitalization process has emerged strongly in the industry, causing an increase of connected sensors and IIoT devices, which produce a great amount of varied data. However, some industrial variables are hard to measure because of its high cost, complex installation mechanisms or non-stop production requirements. These variables could be indirectly estimated based on other related variables available in the process. Data-driven methods would be appropriate for this purpose, modelling real and potentially complex industrial processes. In this paper, a methodology to develop a virtual flow meter for industrial processes is presented. It assumes the impossibility of installing a flow meter in the process, so a non-invasive flow meter is used punctually to measure and capture data for training data-driven methods. Three different methods have been trained to obtain the model function: multiple linear regression (MLR), multilayer perceptron (MLP) and long-short term memory (LSTM).The developed virtual flow meter has been tested on a pilot plant built with real industrial equipment. LSTM method yields the best performance in the flow estimation, providing the lowest MAE and RMSE errors. It is able to consider temporal dependencies, besides modelling the nonlinear nature of industrial processes. |
*** Title, author list and abstract as seen in the Camera-Ready version of the paper that was provided to Conference Committee. Small changes that may have occurred during processing by Springer may not appear in this window.