Fluid classification with integrated flow and pressure sensors using machine learning

This paper describes fluid classification methods using machine learning applied on a microfabricated Coriolis mass flow sensor with integrated pressure sensors. The latter are positioned upstream and downstream of the Coriolis mass flow sensor, which enables the measurement of the viscosity-dependent pressure drop. The Coriolis mass flow sensor itself is particularly sensitive to the mass flow and density of the fluid. Five different liquids (nitrogen, water, isopropanol, ethanol and acetone) are applied to the sensor system in different combinations of mass flow rate, pressure and temperature. For each combination, the raw signals from all sensors are amplified, demodulated, digitized, sampled and stored. Then BiLSTM and CNN neural networks were trained and tested by using train-test split validation and K-fold cross-validation. With both methods, the classification accuracy is determined using a different part of the dataset than for learning. For mass flow rates up to 5 g/h, pressures between 4 bar and 6 bar and temperatures between 288 K and 308 K. BiLSTM performs best with a cross-validated accuracy of 77% up to 100%, dependent on the inclusion of low-flow data.