Resonant sensors for the measurement of rheological properties like density and viscosity are often employed in online process-monitoring applications. Micro-acoustic MEMS devices such as micro-cantilevers and Surface Acoustic Wave devices have been widely used for such measurements. However, due to their scale, these devices measure thin film viscosity and density. Such measurements are often not comparable to macroscopic measurements obtained through conventional devices. Miniaturized cantilever-based devices provide an interesting alternative as they are minimally intrusive, like micro-acoustic sensors, yet measure in bulk rheological domain. However, the interactions between the liquid and the oscillating beam are more complex to model. Such interactions have been previously modeled using classical linear Euler-Bernoulli beam theory or by considering an equivalent lumped elements oscillator such as a Duffing or Van der Pol oscillator. The derived models are subsequently used to relate the liquid's viscosity and density to measurable parameters such as resonance frequency f0 and resonant mode quality factor Q-1. Currently, there are no exact models in the literature for describing the nonlinear vibration of partially immersed beams, nor experimental results providing an understanding of how fluid properties affect the nonlinear vibration characteristics.This work focuses on first establishing empirical relationships between different experimental parameters -such as fluid volume, density, viscosity, length of beam, etc.- and a force-excited, partially immersed beam's resonant response. Then, it describes an ensemble machine learning (ML) based approach that models the nonlinear change in the frequency response of the beam with an increase in excitation amplitude, by measuring the variation in resonance frequency and quality factor of a selected sensitive mode. These measured quantities are subsequently used as features on which ensemble ML models are trained to predict the density and viscosity of the tested fluids. With relatively few (275) training data points, the model can predict viscosity with a 96.65% (R2, 10-fold cross-validation) score and both density and viscosity with an 89.58% (R2, 10-fold cross-validation) score. The impact of this work is to provide a proof of concept for a rheological property sensor that utilizes an ML-based approach for online viscosity and density measurement.
Read
