(470e) Use of Acoustic Emission for Identifying Gas-Liquid Mixing Regime in Agitated Vessels at Different Scales Applying Machine Learning | AIChE

(470e) Use of Acoustic Emission for Identifying Gas-Liquid Mixing Regime in Agitated Vessels at Different Scales Applying Machine Learning


Forte, G. - Presenter, Johnson Matthety
Alberini, F. - Presenter, University of Birmingham
Simmons, M., University of Birmingham
Stitt, H., Johnson Matthey
Use of acoustic emission for identifying gas-liquid mixing regime in agitated vessels at different scales applying machine learning

Giuseppe Forte, Federico Alberini, Mark J. H. Simmons, E. Hugh Stitt

Johnson Matthey Technology Centre, Billingham, TS23 4LB, UK
School of Chemical Engineering, University of Birmingham, Edgbaston, B15 2TT, UK


Operations involving Gas-Liquid agitated vessels are common in the biochemical and chemical industry; ensuring good contact between the two phases is essential to process performance. In this work, Acoustic Emission (AE), using a piezoelectric sensor, was applied to evaluate gas-liquid mixing regime within two-phase (gas-liquid) and three-phase (gas-solid-liquid) mixtures in a 3L stirred tank equipped with a Rushton Turbine and a ring sparger. AE has been previously applied in similar applications in pipelines [1] and to characterise bubble shape and size [2-4]. While moving up through the vessel, gas bubbles will collapse, break or coalesce generating sound waves transmitted to the wall and to the acoustic transmitter. The system was operated at different flow regimes (non-gassed condition, loaded and complete dispersion) achieved varying impeller speed and gas flow rate, with the objective to use features in the acoustic spectrum to univocally identify the different conditions. The methodology was then applied to systems at larger scale, 10, 100 and 1000L to assess performance at manufacturing volumes.


The used apparatus was assembled as for similar works in the literature [5]: a piezoelectric sensor (Vallen Systeme GmbH, Icking, Germany), with resonance frequency of 375 kHz and diameter of 20.3 mm, was attached to the tank with a silicone based vacuum grease. The sensor was connected to a preamplifier (40 dB gain, Vallen Systeme) and an oscilloscope (5243A Pico® Technology Limited) controlled by a laptop was used to record data. Measurements at each condition were taken for 0.2 s with a sampling rate of 750 kHz, ensuring coverage of the resonance frequency of the sensor in the spectrum (Nyquist rule). The sensor was placed on the 3 L Perspex vessel outside wall at a height corresponding to the impeller region.

Machine learning was used to process the acquired data applying logistic regression algorithm. The acquired data in the time domain were preprocessed using Fast Fourier Transform, filtered and selected with respect of their variance, before being fed to the machine learning algorithms.

The dataset was divided into a training set, used to train the machine in recognising features in the spectrum corresponding to each condition, and a testing set, unseen by the machine, used to evaluate accuracy of regime prediction. The training dataset comprised of runs at different regimes in the bi-phasic condition (gas-liquid), while in the test, the system was challenged to recognize the same conditions in the two-phase case and in case of three-phase mixing, where stainless steel particles (3-5 % w/w, density ρs = 8000 kg m-3, size between 0.177 and 0.420 mm) were added to the mixture.


In first instance, for the biphasic case, a dataset composed by 100 spectra was fed to the trained algorithm to be tested. The machine, based on conducted training, assigned to each spectrum a predicted class (ungassed, loading or recirculation). In table 1, the results of the test are reported as number of correctly predicted cases and failures for the biphasic case.

Real condition Correctly predicted Failed to predict
Ungassed 35 5
Loading 26 4
Recirculation 29 1
Global 90 10

The five spectra are wrongly classified as loading instead that ungassed and therefore are counted as misclassified. Other few spectra are not classified correctly by the algorithm, that however achieves an accuracy of 90% on the carried-out study.

In the second part of the study, stainless-steel particles were added at different concentration and the acoustic signal was acquired at the three regimes. The obtained results are summarized in Table 2.

Real condition Correctly predicted Failed to predict
Ungassed 90 10
Loading 66 4
Recirculation 150 10
Global 306 24

In this case the dataset comprised of 330 spectra and the accuracy increases to 92.3%, with an even distribution of errors across the different conditions. The presence of the solids does not affect the classification algorithm that using the training acquired in the bi-phasic condition, it is still able to correctly identify the regime characterising the analysed spectrum. The reason for this can be found in the influence of the solids on different frequencies from the ones characterising the air bubble disruption and impacts.


In this work, Acoustic Emission was used in combination with a machine learning algorithm, logistic regression, to identify gas-liquid mixing condition in a stirred tank. The machine was trained to recognize ungassed, loading and recirculation condition using a training dataset, consisting in the acoustic spectrum acquired at the different regimes in the biphasic condition (gas-liquid). The system was then challenged to recognise such regimes both in two-phase and three-phase condition (obtained adding stainless steel particles), achieving an accuracy equal to and above 90% respectively. The developed methodology is extended to larger scale in order to evaluate physical and practical reasons that would undermine its application in industrial environments. The conducted study aims to propose AE as a potential diagnostic and condition monitoring technique for fluid mixing applications in combination with the use of machine learning algorithms, for applications at large scale for process monitoring.


[1] Addali, A., Al-lababidi, S., Yeung, H., Mba, D., & Khan, F. (2010). Acoustic Emission and Gas-Phase Measurements in Two-Phase Flow. Proc. Institution of Mech. Eng., Part E: Journal of Process Mechanical Engineering, 224(4), 281–290.

[2] Leighton, T. G. The acoustic bubble, 1994 (Academic Press, London).

[3] Manasseh, R. Acoustic sizing of bubbles at moderate to high bubbling rates. In Proc.4th World Conference on Experimental heat transfer, fluid mechanics and thermodynamics, Bruxelles, Belgium, 1997, pp. 943–947.

[4] Al-Masry, W. A., Ali, E. M., and Aqeel, Y. M. Determination of bubble characteristics in bubble columns using statistical analysis of acoustic sound measurements. Inst. Chem. Eng., 2005,83 (A10), 1196–1207.

[5] A. Nordon, R. J. H. Waddell, L. J. Bellamy, A. Gachagan, D. McNab, D. Littlejohn G. Hayward. Monitoring of a heterogeneous reaction by acoustic emission, The Royal Society of Chemistry: the Analyst, vol.129, pp.463-467, 2004.