(178b) The Possible Application of a Single Passive Acoustic Emission Sensor to Identify Different Complex Fluids in a Fully Flooded Pipe

The ability to measure process and product parameters are key for any manufacturing process. Modern sophisticated manufacturing heavily relies on constant measurements in order to guarantee a consistent product quality. Recent trend is the move towards smart manufacturing or more commonly referred as Industry 4.0; a term first introduced by the Association of German Engineers and the German government in 2011, which simplified describes the need to make use of networks and big data in order to meet modern manufacturing and processing needs [9].

One example of a technology delivering high volumes of data and information are acoustic emission sensors (AE sensor).

Such sensors can either be active, meaning they are based on an emitter-receiver system, or passive. Where active acoustic emission sensing measures the change of introduced acoustic waves over the distance from the emitter to the receiver, passive acoustic emission sensors only detect the acoustic emission that the process emits itself. Although active acoustic emission sensors are more present in research and in industry they face several challenges, which include loss of signals due to obstructions, bubbles and distance or simply not reaching the necessary penetration depth [1, 7]. Overall, active acoustic emission sensors prove to give good predictions on factors such as flow rate, degree of gassing or solid content. This technology works well for Newtonian and non-Newtonian fluids [11, 15].

There have been several attempts to make use of acoustic emission data to gain a greater understanding of pipe flow [3, 8, 12], however, work has mainly been focussed on active acoustic emission [2, 5], multiphase flow [8, 14] or water [10, 12, 13].

Passive acoustic emission sensors are used for leak detection in water pipes by employing in-pipe hydrophones [4, 10] or by aiming to recognise acoustic patterns based on the signals of a series of sensors [12]. Other work on passive acoustic emission sensing has mainly focused on multiphase systems [6, 8, 14]. However, no work was done so far on enclosed, fully flooded and single-phased pipe systems.

The work presented makes use of a single passive acoustic emission sensor, which is placed on the outer wall of a 1-inch (ID) stainless steel pipe (L=120 mm) together with a new, self-developed device that makes is possible to receive relevant AE signals. Along with AE signal recording 2D Particle Image Velocimetry (2D PIV) is used in order to visualise the flow fields for different fluids and obstruction types. The 2D PIV chamber used meets the same inner diameter as the stainless steel pipe, hence, is assumes to be representative for the flow fields within it. Four in-pipe obstructions are used on four different test fluids, which all find application in the FMCG sector. The two Newtonian fluids used are water and glycerol (CAS Number: 56-81-5), whereas Carbopol® (CAS Number 9003-01-4) and Sodium carboxymethyl cellulose (CAS Number 9004-32-4) examples of a shear-thinning fluid. This choice of fluid enables the investigation of laminar and turbulent flow as well as the effectiveness of passive AE on complex fluids. Further pressure drop Δp_v is determined in each experiment, showing that different obstructions pump settings and test fluids lead to different readings.

The test rig consists of a water recirculation system powered by an I KA-5 132SSS1 centrifugal pump (Alfa Laval, Sweden). The in-pipe obstructions are all designed in AutoCad 2018 (Autodesk Inc., USA) and extruded by a FlashForge Dreamer 3D printer (Zhejiang Flashforge 3D Technology Co., Ltd., China). Obstruction types can be slotted into the pipe/ 2D PIV chamber and are all significantly geometrically different. Geometries are a cone, cross (four wedges meeting in the central focal point), three triangular aligned holes and a semicircle. The TSI 2D Particle Image Velocimetry system (TSI Inc, USA) uses a green 532 nm Nd-Yag laser (Litron Nano PIV) pulsing at 7 Hz and is synchronised to a single TSI Power View 4 megapixels (2048 x 2048 pixels) 12-bit CCD camera. The CCD camera is connected to another synchroniser (TSI 610035) which is attached to a desktop PC. The 2D PIV system is controlled by TSI Insight 4G software. For cases in turbulent flow 500 images are taken, whereas for cases of laminar flow 100 pictures. The area is set to 32 x 32 pixels. The pictures are afterwards combined and averaged and debugged in order to determine the flow field. The final pictures are visualised with the help of Tecplot 360 software (Constellation Software, Canada).

Acoustic emission signals are captured with a piezoelectric VS375-M sensor (Vallen Systeme GmbH, Germany) that is linked to a 2.5 kHz to 2.4 MHz (10 Vpp) AEP5H preamplifier (Vallen Systeme GmbH, Germany) along with a DCPL2 decoupling unit (Vallen Systeme GmbH, Germany), a PicoScope 5000 Series oscilloscope (Pico Technology Ltd, UK) and a personal computer using PicoScope version 6.13.15 software (Pico Technology Ltd, UK). 100 buffer, each of a length of 500 ms, a resolution of 16-bit and an amplitude of maximum ±1 V. The sampling number set to 600 kS to ensure that the sampling frequency is at least twice the number of the resonance frequency [16]. The choice of 500 ms is justified as a 10 s buffer is divided that long until the lowest recording time reached where Fast Fourier Transform (FFT) spectra are still visually identical.

The recorded acoustic emissions are pre-processed aiming to reduce the size of input data to a level that it can be used for supervised machine learning (ML) on the given computational power of a 8 GB memory. In a first step, the noise must be removed. As environmental noise those frequencies below 4 kHz are considered and the FFT spectrum is extracted. In addition, positive and negative infinitive values are filtered and replaced by the value ±1. The last reduction step is the selection of only the 5,000 FFT values with the largest relative variance.

To make the data available for the Matlab R2018a (MathWorks Inc, USA) Classification Learner Application (CLA) feature scaling and mean normalisation was applied to the reduced FFT spectrum. The necessity of this operation is to have features with a comparable range of values.

Principal Component Analysis (15 components) is enabled for the CLA app. The goal of the PCA is to project the data points of the matrix into an n-dimensional subspace in such a way that as little information as possible is lost and existing redundancy is summarised in the form of correlation in the data points. This enhances the supervised ML performance, contributing to a better support vector delimitation. The dataset for the Machine Learning is further split into in a training (60%), optimisation (20%) and a test dataset (20%). The test dataset is given to the supervised classifiers (e.g. KNN, Decision Tree, SVM) whilst the second dataset is not fed into the CLA to evaluate the accuracy of the algorithms.

For the acoustic data, higher flow rates lead to higher peak values and different pipe obstacles to different FFT spectra. Also, different fluids lead to different amplitudes in acoustic signals with highest peaks for water. Highest accurate predictions in supervised ML are attained with quadratic SVM with accuracy levels being in the region of 95%. This applied as well as to different flow rates, fluids and obstruction types. In addition, 2D PIV supports differences as different types of obstacles lead to different flow regimes and increases in flow rate lead to increases in velocity magnitude. Besides, different fluid types show different flow fields.

Summarising, the application of a single passive acoustic emission sensor linked to supervised ML is proven to give accurate predictions on flow rates, obstructions and fluids. The technology works for Newtonian and Complex Fluids. This may help in the future to predict flow regimes, improve process design and operation conditions. More importantly, the system for single-phased fluid flow, which is not documented in current literature.

Future research can focus on refining the data processing or making use of neural networks for broaden application fields. Also, for each case single algorithms are created. To enhance user friendliness those need tying together or the creation of some sort of cascade algorithm and the creation of a user-friendly interface.

Acknowledgement

The authors wish to express their gratitude for the support received from Philip W Harris and Robert W Sharpe in the engineering workshop at the University of Birmingham. Further thanks goes to Dr Thomas Mills providing necessary training and enabling access to the 3D printer system.

Funding Note

The School of Chemical Engineering of the University of Birmingham supported this work.

The authors declare no conflict of interest.

Funding Note

Please note: A reference list will exceed the total word count. Hence, please advise how this can be supplied as well. Thanks. Daniel Hefft