Optical inline analysis and monitoring of particle size and shape distributionsformultipleapplications:Scientificandindustrialrelevance

J?rn Emmerich,Qiao Tang,Yundong Wang,Peter Neubauer,Stefan Junne,Sebastian Maa? *

Keywords:Particle Multiphase flow Algorithm Inline Photo-optical Concentration

A B S T R A C T Particles occur inalmost all processesinchemicaland lifesciences.The particlesizeand shapeinfluencetheprocessperformanceandproductquality,andinturntheyareinfluencedbytheflowbehavioroftheparticlesduring production.Monitoring and controlling such characteristics in multiphase systems to obtain sufficient qualities willgreatlyfacilitatetheachievementofreproducibleanddefineddistributions.Sofar,obtainingthisinformation inline has been challenging,because existing instruments lack measurement precision,being unable to process overlapping signals from different particle phases in highly concentrated multiphase systems.However,recent advances in photo-optics made it possible to monitor such features(particle size distribution(PSD),aspect ratio and particle concentration)with advanced image analysis(IA)in real-time.New analysis workflows as well as single feature extractions from the images using multiple image analysis algorithms allowed the precise real-time measurements of size,shape and concentration of particle collectives even separated from each other inthreephasesystems.Theperformances,advantagesanddrawbackswithothernon-photo-opticalmethodsfor assessing the particle size distribution are compared and discussed.

1.Introduction

The implementation of process analytical technology(PAT)in the chemical production has a tradition of a century.Early examples are the Siemens&Halske flow type conductivity meter to measure carbon dioxide[1].However,despite the long history and extensive developments in the area of PAT,the particle size has not become a standard operation parameter for monitoring.Although such information is usually important,monitoring of particle sizes and shapes is not easy.Essentially,particles are a portion of something suspended in something else like a solid in a liquid(suspension),a gas in a liquid(bubblyflow or foams),liquid in a liquid(emulsions),solid or liquid in air(sprays).Many properties of particles are influencing for the process,e.g.their size,their size distribution,their shape,their surface structure,their fragility,their charge,which impacts their stability,and their porosity/microstructure,respectively.

Therefore,measuring the particle properties provides additional information in the concerned processes.There exist numerous products,in whichparticle properties are important,like polystyrene,cement,inhalers,batteries,sunscreen,sauces,flavorings and plastic.Liquid-liquid systemsconsistoftwo phasesofimmisciblefluids.Current improvements and optimization of many processes rely on a proper understanding of the behavior of such liquid-liquid systems.Another target of optimization for catalyst suppliers and users of fluidized-bed,slurry-batch/CSTR(continuous stirred tank reactor)or bubble-column reactors is catalyst attrition.Catalyst attrition is the process by which catalystparticlesslowlybreakapartduetocollisionwithotherparticles,stirring devices or reactor walls.This presents a challenge to produce catalyst material with sufficient mechanical strength under realistic reaction conditions to achieve the required lifetimes.Additionally,not only the size of the catalyst,but also the bubble size distribution in such multiphase reactors is of high interest.For an exact prediction of the heat and mass transfer as well as reaction kinetics,an exact knowledge of the interfacial area is required.The quantification of bubble size distributions is a pre-requisite to achieve this.

Among the most promising methods for inline bubble,drop or particle size measurements are photo-optical captures coupled with image analysis.The term‘inline'is used in the sense that the measurement instruments are directly installed inside the multiphase system(seeFig.1).Ithasbeenshowntobemorereliablethanmanyalternative measurement techniques[2-5].A review about market available technologies is presented as well as a number of recent case studies using a photo-optical system,that is comparable to that of the various other existing photo-optical devices for particle characterization.Image analysis tools combined with direct and inline data acquisition open new opportunities for academia to gain fresh insights into processes as well as new ways for optimization and process transparency for industrial production plants.

Fig.1.Categorization of measurement techniques and terminology “offline”,“online”and “inline”for particle size and shape analysis[8,9].

Existing technologies and its applications are reviewed extensively in the literature.A number of applications of photo-optical probes are presented in the following sections.The complexity of the applications increases from a pure liquid-liquid system,over liquid-liquid-gas to a high pressure,high temperature application and a crystallization in an air loop reactor.Applications in the area of cell suspension are thoroughly discussed as well.

2.State of the Art

There exist a broad variety of instruments for inline particle characterization,rangingfromtechnologiesbasedonacoustictoimpedanceor laser back scattering.These methods,although applied in various processes,are not able to provide a capture of particles,which allows for a manual discrimination[6].This is a general drawback whenever the size of a particle is quantified based on an indirect method.Nevertheless,although any photo-optical determination has the advantage of providing an image of each individual particle,the major challenge for the industrial application of photo-optical instrumentation is the dif ficulty in processing the acquired data reliably and automatically.Many efforts have been made in obtaining real-time data from photo-optical instruments in the field of particle characterization[6].The inline and offline particle analysis methods that are commonly employed for particle measurements will be described in Section 2.1.Section 2.2 subsequently provides an overview of experimental data and also gives an overview of commercially available laser based and photo-optical sensors[1,7].

2.1.Particle size measurement as process analytical technology(PAT)

The numerous measurement techniques for the detection of multiphase system properties are based on various physical principles and have corresponding advantages and disadvantages in their operation,system specific applicability and accuracy.A subdivision of the measurement techniques can be made on the basis of their process-related temporal and spatial operation,e.g.real-time methods that deliver data in a process-relevant time window are opposed over offline methods.A further distinction is made between measurement techniques that record measurement data inline(inside the process),or by external sampling.To characterize the contact between measuring technique and the system to be measured,the classification into invasive(intrusive)and non-invasive(non-intrusive)methods is made.A corresponding classification is given by Gnotke[8],see also Fig.1.

The term offline is used for manual sample taking and analysis.The sample needs to be prepared(i.e.diluted,mixed)for the actual measurement and the time between sampling and results presentation exceeds normally the time of 60 min.Results for online measurements need to be available within shorter time(＜10 min.).Such measurements are directlyattached to theprocessvia a by-pass with automated sampletaking.Thetimein which informationaboutprocessormaterial properties is obtained needs to be shorter than the time in which these properties may change.‘Inline'measurements have a permanent direct contact with the process stream and provide continuous data from the process.The time between data acquisition and result presentation is comparable to online measurements.

Intrusive offline,online and inline measurements may potentially disturb the process.Additionally,sampling and sample preparation accompanied with deviations from the process environment may cause a change in particle interactions,which results in measurement errors,especially in unstable dispersions[10].It is therefore desirable to minimizethe chance of particle deformation or the change in particletrajectories to a large extent by an appropriate placement of the invasive instrument into the system[11].

For a further subdivision of inline measurement techniques in liquid/liquid dispersions,Maa? et al.[3]uses the underlying physical principle of measurement data acquisition.They name the four major classes of acoustic,laser-based,photo-optical,and Coulter-based systems[12].From the multitude of measuring techniques for the determination of the particle size in dispersed systems,a number of techniques are described in the following.

Laser-based,photo-opticalandacousticmethodsareregardedasthe most relevant measuring techniques for determining transient particle size distributions.For bubble size distribution measurements photography[13,14],the capillary suction probe(CSP)[15,16]and the PDA[17,18]are the most widely used techniques.Fig.2(left)shows the CSP measurement technique developed by Barigou&Greaves[19].It is a short,thick-walled glass capillary with 0.39 mm inside diameter and funnel-shaped inlet.Due to the housing of this precision capillary in a stainless-steel tube with suction and two centrally located measuring points,eachconsistingofanLED(lightemittingdiode)andaphototransistor,thereforealsonamedphotoelectricsuctionprobe.Thevolumesof theparticles sucked in ata constantspeed,whichare elongated,behave proportionally to the length of these cylindrical plugs and are detected by the two light barriers.

An example for a non-invasive measurement technique recording the spatial particle size distribution is given by the short-time holography(STH).With this technique all optical information of the light,which is scattered and reflected at thephase boundary from the surface of a particle can be recorded.Objects that exceed approximately ten timesthesizeofthewavelengthfromthelaserlightsourcearedetected and visualized very sharply due to the short exposure time.The one shown in Fig.2(middle)by Ilchenko et al.[20]was used to study the bubble size distribution in a cylindrical stirred tank.

Anotherprinciplemeasuringparticlesizedistributionaswellasparticle velocity analysis is used by the Inline Particle Probe 30(IPP 30,see Fig.2,right).Itisbasedontheprincipleofamodifiedfiberopticalspatialfiltering method,which is an extension of the spatial filtering velocimetry(SFV)by means of the fiber optical spot scanning(FSS).The SFV refers to the method of speed measurement of an object by viewing it through a spatial filter,which is placed in front of a receiver(photodetector).The determined speed is used to calculate the particle size by the FSS,which detects the silhouette of a moving particle by obscuration.

2.1.1.Acoustic and Coulter-based systems

Schlüter[22]provides a comprehensive overview of currently applied measurement methods for determining process relevant parameters of continuous and dispersed phases.In order to record the disperse phase properties(particle size,velocity,distribution),the methods are classified depending on the optical properties of the system to be measured:(i)opaque media with high disperse phase proportions,and(ii)transparent media with low disperse phase fractions.The needle probe,the ultrasound bladder analyzer(UBA),the endoscope or fiber-optic instruments and the phase-Doppler anemometer(PDA)for the measurement of particle size addressed in[22]are relevant in this work.The UBA uses the relationship between scattering of the ultrasonic signal and curvature of the phase interface immediately in frontoftheprobetomeasurethebubblesizedistribution.Itcanalsodetermine the bubble velocity by means of the Doppler shift of the ultrasonic burst and the local gas content based on the number of particles per time.The phase-Doppler anemometer(PDA)can simultaneously measure the diameter of spherical particles and their velocity.Both methodologies provide meaningful data in relatively transparent media and rather low disperse phase fractions.Coulter based systems rely on the impedance measurement in a capillary,through which a dispersed electrolyte solution flows.When a particle passes through the measuring gap,it displaces the electrolytic solution by its volume,which is proportional to the measured resistance,depending on conductivities of the continuous and the dispersed phase.Such systems are widely used for offline measurements and have not been used inline,so far.

2.1.2.Laser based systems

Laser-based measuring systems can be classified into three main groups by their physical measuring principle:Fraunhofer diffraction,local filter technology and laser backscattering.Laser backscattering is the most widely used inline technique[23].Maa? et al.[3]tested three laser-based probes in a concentrated toluene/water system(dispersed phase fraction ?=0.2)as shown in Fig.3 for droplet size measurement in a stirred tank.All probes work according to the laser backscatter principle and can be used for intrusive measurements.

The two-dimensional optical reflectance measurement(2D-ORM)technique(see Fig.3,left)uses the reflection of an intense laser beam to detect the chord lengths of emulsion droplets in close proximity to the window at the end of the probe tip.The focused laser beam can be controlled in a very thin 2D layer of the emulsion,in which the collimatedlaserrotatesataveryhighspeedinacircularpath.Themeasured arc lengths of the particles are converted into a chord length distribution(CLD).From this,the particle size distribution(PSD)can be obtained by further mathematical transformations.The 2D-ORM is especially suitable for very dense emulsions or suspensions,as the particle sizes are detected by reflection in a field very close to the probe tip[3].

The working principle of the forward-backward ratio(FBR)fiber sensor,seeFig.3(center),isbasedontheanalysisofthespatialintensity patternofforwardandbackwardlightbackscatter,whichischaracteristic for particles outside the Rayleigh scattering range(that are particles with a radius less than a tenth of the wavelength of the incident light).For larger diameters,an increasing asymmetry between forward and backward scattering is observed.From the ratio of the light intensities scattered in both directions,the mean diameter of a particle group can bedetermined.Depending on thesensorand the lightsourceused,particles in a size range from 50 nm to 200 μm can be measured[25].The three optical fibers,one for illumination and two for the detection of scattered light,are arranged at an angle of 60°to each other[3].

Fig.2.Photoelectric suction probe:CSP metrology(left)[19];components of the experimental design and operating principle of the optical measurement method STH(middle)[20];principle of IPP 30[21].

Fig.3.Laser-based probes for measuring the particle size distribution:2D-ORM sensor(left),FBR sensor(center)[3],principle of the FBRM probe(right)[24].

Themostcommonlyusedlaserbasedinlinemeasurementtechnique is the focused beam reflectance measurement(FBRM),see Fig.3(right)[7,26].The optical rays from this cylindrical probe rotate at high speed andfocusthelaserbeamnearthesapphireglasswindow.Particlespassingtheprobetip scatterthe laser lightback so that thechordlengthof a particle can be calculated by multiplying the sampling time by the rotationalspeed.Basedonthisprinciple,severalthousandchordlengthscan be recorded every second[24].

2.1.3.Photo-optical based systems

Based on literature,it becomes obvious that laser and photo-optical systems offer the most promising results for detecting disperse phase properties.Both methodologies provideaverage particle sizes and complete distribution functions.Photo-optical systems with image analysis have become a standard method for offline particle quantification[2,4].Due to the immense time required for manual image analysis,however,it has rarely been suitable for online surveillance yet.

Over the last decades the interest,as an alternative to CSP,optical methodshasincreased significantly.Thecause lies intherapid andcontinuous development of digital cameras,lasers and automatic image processing(IP),which has led to a considerable high speed of the quantitative optical measurements.High-speed digital cameras allow photography to capture both the size of particles and their projected shape providing a distinct advantage over other measurement techniques,such as capturing CLDs.Captured images of particles in dispersions are a trusted measurement for the quantitative determination of particle size and shape distributions and offer a reliable standard procedure for validating the results of other measuring devices[27].

A number of photo-optical devices for particle characterization have emerged in the last 20 years,that all face the same variety of applications.Pacek et al.[4,28]were one of the first to use a microscope combined with a video camera for the purpose of monitoring particles(droplets,bubbles andthelike).Their techniquecombined themagni fication power of a microscope with the image acquisition capability of a video camera.The research group from the University of Birmingham was the first to apply this technique to stirred liquid-liquid systems.ThecomparisoncarriedoutbyPaceketal.betweentheusedvideotechnique and standard technologies from that age(capillary suction method,Pilhofer method[29])did set the foundation for the triumph of photo-optical methods over those by now almost being forgotten.Pacek and Nienow[5]described the used hardware in detail:a conventional 700-line high quality video camera that took images through a stereo microscope from outside the vessel near the wall(max.8 mm)or inside the vessel through a light tube.Both methods are intrusive since lighting is provided from behind the focal plane by a strobe light guided through a thick fiber optic tube.The light was edited to be synchronized with the camera.Pictures were taken at a frequency of 50 frames per second(fps)and scanned for non-overlapping drops by an object recognition software.It required human intervention for correction and to categorize drops pictured inside of larger drops.Diameters of 25 μm up to a few millimeters can be recognized.The fast and continuous data acquisition and the flexible positioning of the light tube are advantageous.The intrusion of this tube and the fiber optic tube as well as the outdated hardware is disadvantageous.Presumably,the technique was expanded to a digital video camera and improved image recognition[30,31].Wang et al.[32]describe a telecentric optical system with reduced perspective distortion and LED illumination.Generally,smaller distortion ratios lead to more realistic projections of photographed particles on the image sensor.Integrated LED illumination is an alternative for fiber optical pathways,although with limited photon flux density for short light pulses and temperature range in operation.Xiao et al.[33]address the general challenge of poor light conditions in particulate systems with a particle scattering photography approach.The study is focusing on solid particles with a size range between 0.5 and 1.5 μm.Further developments in inline photo-optical sensors and their diversification over the last decade are described in the following section.

2.2.Overview of experimental data using inline particle size and shape analyzer

Maa? et al.[3]compared four inline probes to measure drop size and bubble size distributions.The often used FBRM,2D-ORM,IPP 30,FBR and an endoscope with image processing have been tested under the same process conditions.By comparison with the self-referential results of an inline endoscope using image analysis with semi-automatic particle recognition(500 drops·h?1),the challenging task of measuring the drop size distributions in liquid/liquid systems could not be satisfactorily addressed by any of the laser probes.The dependence of the drop size distribution on the power input ε gave reasonable results for both the FBRM and the 2D ORM.For these two laser-based probes,the same proportionality of the Sauter diameter d32over ε could be determined as for theinlineendoscope.However,laserprobesdidnotprovideplausibleresults for d32at constant stirrer rotation frequency n over time t.

Based on these results,the general question arises as to whether the principle of laser backscattering is suitable for investigating drop size distributions in liquid/liquid dispersions.The smooth surface of the drops causes great errors in the measurement because the laser light is scattered back by one drop in three separate beams.The measured chord length is accordingly too small.To further analyze this effect,Maa? et al.[3]artificially rendered a rough surface by the addition of titanium dioxide microparticles(TiO2).The examination of this manipulated dispersed system provided more plausible results by the FBRM.However,in accordance with[34],the absolute values obtained by the FBRM are significantly smaller than the values of the endoscope measurements with image analysis.The reason for this is FBRM measures a chord length distribution(CLD)with reasonable efforts to transfer theCLDintoaPSD[35],whichisrelatedtothesizeandshapeoftheparticles in dispersions.The IPP probe system has been tested successfully in a fluidized bed application[36],but shows disadvantages in liquid dispersions,especially when flow directions are inconsistent.

Todtenhaupt[37]performed investigations in a bubble column and a pressurized stirred tank by means of a suction probe with a light barrier.Pilhofer&Miller[29]provide a detailed analysis of a photoelectric probe and compared the measuring accuracy of a single and doublepass probe.In their investigations,Barigou&Greaves[15]included local bubble size distributions at 63 measurement points in the high shearrate stirrerregion.They compared theverticaland radialdistributions of the Sauter diameter d32and provided a strong inequality of these due to the inhomogeneous turbulence.Alves et al.[38]used the CSP technique to study a stirred tank with a two-stage disk stirrer arrangement.A disadvantage of the suction probes is the limitation of the correct size of the particles,which are to be stretched out in length withinthecapillary.Ifthesizerangeistoowide,differentprobes ofseveral capillary diameters should be used,as larger particles break up.Ilchenko et al.[20]positioned the STH in a rectangular vessel to avoid optical distortions.Glycerin was filled in the gap,which has the same refractive index as the container glass.The pulsed emitted light(wavelength of λ=694 nm)illuminates a holographic image plate for a period of time of t=30 ns through a laser beam passing through thestirredtankandareferencebeam.Thisresultsinastorageoftheentire optical information of the measuring section on the coating of the plates.After illumination,the holographic image plates are developed in a similarwaytoa black-and-white photographand henceforth called a hologram.Following this,the original test volume can be reconstructed by a continuous wave laser(HeNe,helium neon)directed at the hologram,which simulates the reference beam during illumination,andcapturedbycamerashotsatvariouspositionsintheformofa2Dseries of images.Mayinger&Feldmann[39]used an extended stereoscopic setup with a simultaneous recording of two holograms oriented perpendiculartoeachother.By matchingthetwoholograms,both bubble shape and position can be determined exactly.However,the enormous advantages of this technique of taking a snapshot of the reactor in order to be able to reconstruct it three-dimensionally in hindsight are not available on an industrial scale.The necessary preparations to ensure the optical accessibility as well as the development and reconstruction of the holograms allow only a use for research purposes on a laboratory scale.The main challenges of this methodology are a comparably complex setup that is not widely applicable.

Photo-optical inline instruments which have been used for cell concentration measurement are not in the scope of this study[40].However,microscopy is a common tool to observe the cell and microbe size and shape.The previously described advances in the field of inline photo-optical monitoring are also of great importance for monitoring such features directly in bioprocesses.In many cases,a direct relation between cell morphology and its function exists[41],which enables to monitor many parameters indirectly with the application of inline photo-optics.Nevertheless,several challenges appear when measuring directlyintheculturebroth:liquid(media),solid(cells)andairbubbles in aerated systems have to be distinguished from each other.Besides,particular cell features,which are of interest,have a size smaller than 1 μm,which makes monitoring complicated.Furthermore,the usually applied complex media possess many insoluble particles,which demands a sophisticated image analysis and absorbs much light.The high cell densities obtained today create many overlapping particles,which cannot be easily identified.

Despite all these challenges,monitoring of life cells has a great potential and is therefore already conducted for many applications,although not often inline due to a lack of suitable devices.First developments towards probes,which are capable to measure in line,were coupled without mechanical sampling[42,47].Advances in automated image detection led to the feasibility to identify overlapping cells,which made a dilution partly unnecessary.Such techniques were firstly applied to animal cells and their viability determination[43].In subsequent studies,the concentration of yeast cells and their size during osmotic stress responses were monitored with inline microscopy[44].Up to now,it became possible to monitor the cell concentration until and beyond 80 g·L?1of dry weight in the case of Pichia pastoris[40],and in the case of the smaller Escherichia coli cells up to a concentration of 70 g·L?1[45].However,no single-cell analysis is feasible yet under those high cell concentrations.Other descriptions about the application of inline microscopes directly in a bioreactor or a bypass were summarized elsewhere[42,46].Immersion lenses were applied to broaden the applicable range of concentrations for feasible measurements at same microscopy settings[47].More and more applications are described,in which such microscopes are applied to determine special morphological features and their relation to physiological features[44,48-50].

Single-cell based photo-optical analysis was conducted especially for larger cells since quite some time.If the cellular metabolism is fast likeinmanybacterialcultivationsoralsoatyeastcultures,arapiddetection of changes becomes necessary.By-pass approaches might alter the metabolism after an exposure to artificial growth conditions(e.g.poor oxygen supply and mixing).This restricts the application rather to slower growing cultures,like cell lines.Besides their importance,their larger size is one of the main reasons why new methodologies are oftenfirstlyappliedtocelllinesratherthanmicrobes.Oneofthosetechniques is single-cell lineage tracking analysis,which was performed by live cell imaging,tracking of every single cell,and determining the fate of individual cells[51].It was used to assess the heterogeneity within a HeLa cell line.In parallel to DNA-based single-cell technologies to identify such heterogeneities[52],photo-optical measurements can provide similar information.Such information can be connected with product synthesis,e.g.if fluorescence microscopy is applicable,while the product provides a concentration-dependent signal for the measurement device[53,54].Although this technique is not new to couple a fluorescence microscope to a process,measurements inside a reactor have not been applied yet and the full potential for process monitoring and control is not used so far.If such methods were coupled to a robust method for cell selection[55],eventually in a sample line,a faster process and product quality control,and on-site process optimization becomes feasible.

Inline sensors are supposed to be applicable in different process environments.The areas of applications are extremely diverse so that there is practically no inline sensor available which is able to cover all given circumstances.Table 1 provides an overview of the main characteristicsof commercially availableinlineprobes.Examplesof laser techniques(e.g.IPP[21]as SFV,FBRM[34]and 3D-ORM[56])or acoustic techniques(e.g.OPUS[57]as SONAR)and alternative video microscopes(e.g.,PVM[34,58],Pixcope[59])are presented.

The Pixcope measurement technology is based on direct optical imaging and depending on the measurement setup,the suspension flows either through a narrow gap in the probe head or flow through cell.Images of the suspension are captured using a transmission setup.Real-time image analysis detects dispersed phase objects in the images and measures their properties including size and shape.Measurement technology based on direct imaging is able to capture detailed particle morphology.Based on the size,shape and color measured for each object,the objects can also be classified.The various inline microscopes from other leading producers(EnviroCam,PetroCam,and ISPV(in situparticle viewer))are listed combined as IVM(inline video microscopy)because of their similar technology and applicability.

Table 1 Overview of invasive measurement instruments for particlesizeand shape analysis

3.Image Capturing and Image Processing Methods

Chemical,pharmaceutical as well as food production processes are sensitive,highly regulated and a substantial foundation of the benefits in our daily life.Reliable data about the particles should be acquired in real-time,without influencing the process or its properties for the purpose of process control and optimization.The measurements have to work for rough process conditions(high or low pH,temperature as well as pressure),for low and high dispersed phasefractions,differentiating all different phases inside the processes.

The aim is to replace sample taking and sample preparation such as dilution by measuring directly within the process stream—may it be a stirred tank,a pressurized reactor,a pipeline,or whatsoever.The current limitations of the presented technologies(see Table 1)become even wider satisfying industrial production related demands like ATEX(atmosphères explosibles)approval,21 CFR(title 21 of the code of regulations)compliance or 3A sanitary standards.

The benefits of competing inline technologies are given by Maa? et al.[3].As for anymeasurement technique,there are some limitations.Firstandforemost,theimagequalityneedstobeaffirmed.Theparticles/bubbles/droplets of interest need to be visible with a certain minimum of contrast.The boundaries of the particles should be visible/detectable,at least partially,as for example for overlapping droplets.Fluid dynamics,viscosity,and light absorption also play a crucial role.

As discussed above,theISOstandards aswellasdiversepublications do favor the photo-optical methods over others.As imaging is a direct method,theworking principle as well as theresults in the same system are equal.That was proven in an example experiment carried out with two different photo-optical probes(PVM[34]and SOPAT[6,60],www.sopat.de).The results are shown in Fig.4,which do show drop size distributions of toluene droplets in water,measured at the same point in time.The images have been analyzed manually(see legend in Fig.4)to analyze the reliability of the method.Furthermore,automated image analysis software was used for the same set of images.The three distributions are,as expected,very similar and show the reliability of the direct measurement method of image-based particle sizing.

One of the commercially available systems was used in further studies to review the applicability of the method to various industrial and research applications.

3.1.Hardware of inline photo-optical particle sizing systems

The general working principles of the hardware of inline photooptical instrumentation have been patented by a number of manufacturers like Olympus,Sumitomo Electric Industries and J.M.Canty in the last decades(see also European patents EP 0343558 A1,EP 1619532 A1;US American patents US 5730701 A,US5619043A and[44]).Such instruments also work in combination with an automated image analysis(more details in Section 3.2)that is able to extract the relevant particle information from the image data[61].Recently,the company SOPAT(www.sopat.de)combined and commercialized individualized hardware with image analysis tools.Different inline probes are available with several magnifications to cover a particle measurement range from around 500 nm to 35,000 μm.The specific magnification must correspond to the anticipated particle size range in order to perform the measurement task(see also setups in Table 2).

Photo-optical measurement systems include the probe with endoscopeor microscopeoptics,acamera,a measuringPC for control,a stroboscope for generating the flash and various transmission cables.The camera used for experimental data presented in this study is connected to the lens system via an adapter(i.e.C-mount)and can shoot up to 19 fps(frames per second)with 2752×2200 pixels(6 megapixels)using a Gigabit Ethernet port.The two-dimensional image of the measuring volume projected onto the large CCD(charge-coupled device)chip(124.8 mm2)is recorded as an image with an information content of 8 bits per pixel(256 grey values)and transmitted via the Ethernet cable to the measuring PC.Color versions of the cameras are also available.On the measurement PC data storage as well as the data processing with following experimental examinations through the specially developed measurement software is performed.

Fig.4.Comparison of drop size distribution(cumulative number distribution)fromtwo different measurements ina toluene/water system.Two photo-opticalprobes doshownearly the same results.

An endoscope can be described as a tubular instrument with an optical system of objective,rod lenses and eyepiece including a lighting device,which is mostly based on the transmission of cold light[62].The endoscope performs the task of achieving a difficult position by its elongated shape,transporting the light(necessary for illumination of the measuring volume)those rays are influenced by the optical processes in the system(transmission,reflection,refraction,diffraction),as well as transporting them to the camera in order to project them onto the chip.

A very important aspect for the realization of a good image acquisition is the illumination.For the deployment in various applications photo-optical systems must get along with the variability for intensity,homogeneity,spectral range and the polarization properties.The qualityoftheilluminationsignificantlydeterminesthesharpness,whichdepends on the aperture and thus the incidence of light,and the contrast of an image.To ensure sharp images even for highly fast-moving particles a stroboscope with a xenon flash lamp is often used.It generates aflash of light with adjustable intensity and a maximum energy of 2.5 J.The flash duration is normally around 2 to 8 μs.The response of the stroboscopetotheflashsequencecontrolisaffectedbyanexternalelectrical pulse.

For illumination several options are applicable(see Fig.5).The simplest one is to take a dark field image by emitting the light rays into the measuring volume and detecting the light returned by the particles(reflexion).Furthermore,a reflection surface,arranged parallel to the viewing window,can beattachedbymeans of a correspondingadapter.Due to the corresponding reflection material and the distance to the viewing window both bright(i.e.white Teflon)and dark field recordings(i.e.rhodium mirror)can be achieved.A representation of thisadapter can befoundin thefollowingFig.5.A so-called U-Lightcan also be used to realize a bright field illumination.It is a bundle of optical fibers,which is housed in a steel tube,parallel to the endoscope with a 180°deflection for direct emission of light on the viewing window.The light emitted by the stroboscope is in this case introduced through the fiber optic cable directly into the U-Light(transmission).For more technical details,see Panckow et al.[6],who describes not only a brief technical description,but also more references and application examples.

Table 2 Overview of applications of the photo-optical SOPAT system

Fig.5.Different illumination methods applicable for inline photo-optical devices.From left to right:a)reflection(light from the top),b)direct and c)diffuse transflection by the use of a reflection adapter and d)transmission with the illumination from the back.

The different illumination methods create individual optical scenes in the measuring volume.The images in Fig.5 show the same agitated air/water system with 0.012%PVA(polyvinyl alcohol)(Alcotex B72)as a surfactant.In this system the reflection(a)of the air bubbles is very complex with partial oversaturation of pixels on the image sensor.Generally,oversaturation of pixels results in a loss of information and is due to unfavorable light conditions.Direct transflection(b)provides a higher contrast compared to reflection(a).However,in systems with particles of high refractive index(i.e.aerated systems)direct transflection often leads to oversaturated regions on the image sensor.Withthetransmission(d)methodparticlesappearoftenasdarkobjects with bright background.Transmission(d)and direct transflection(b)methods are supportive in very concentrated particulate systems.In this example(see Fig.5)diffuse transflection(c)provides the most homogeneously illuminated image sensor.The various illumination methods inherently lead to different particle visualizations due to the distinct paths the light transmits to the image sensor.Under this assumption the size range as well as the size distribution will be different using different illumination methods in the exact same system.Often certified reference materials(see also ISO guide 35:2017,[63])are needed to calibrate the photo-optical devices in order to secure a traceable route to size and quantity of the size distributions they publish.

3.2.Image analysis software

Innumerable algorithms have been developed in the field of particle detection(more precisely,particle instance segmentation)over the years.Mostalgorithmsaredesignedinsuchawaythattheycanonlydetect special object shapes(circular,elliptical,vessel-like,etc.)and often only work well under certain boundary conditions with little flexibility,e.g.,non-overlappingobjects,objectsand/orobjectcontoursmustbesalient,and image intensity statistics should not vary too much.Except from the detection performance of an algorithm its speed may be important,especially in real-time surveillance of processes where intervention might be required.In such a case it may even be desirable to reducethedetectionperformanceforamoreperformantalgorithm.Objects thatare salientandnon-overlappingcanusuallyeasilybedetected by simple Image Processing(IP)approaches.Castleman[64]gives a good overview of available IP methods.A typical algorithm may be based on intensity thresholding or double thresholding in conjunction with filtering methods to enhance objects(contours).Morphological imageoperationscanbeappliedtocloseandfillfoundobjectsandtoremove undesirable small ones.If the produced result contains False Positives(FP)anadditionalclassifierthattakesfeaturesfromtheregionofa segmented object can be used to sort out FP.By means of pattern matching and other computer vision approaches(e.g.,different Hough Transforms),simple geometric objects,like circles and ellipses,or objects of fixed shape can often be well detected also if they overlap or if they are not so salient.Numerous circle detection approaches have been developed and a few of them have been recently compared in a study about the bubble detectors and size distribution estimators[65,66].

Brás et al.[67]developed a very specific image analysis for one liquid-liquid system.The software gave reliable results but was not applicable to other systems,and it gave only poor results when used for different liquids.Recently,Wang et al.applied the commercial software ImagePro Plus[32]for an automated analysis of their particle images.The authors do explain not only the usability,but also the necessity of manual readjustments for dispersions with medium or even high concentrations,as the software fails to recognize the particles separately from each other if they do overlap on the images.

Morerecently,artificialdeepneuralnetworksshowedthattheirprediction performance outperforms classical approaches in a number offields,e.g.,biological cell instance segmentation.A trained net can befine-tuned with training samples of new objects,thus to detect new shapes,whereas in a classical approach a human being would have to extend its handmade algorithm,which is mostly a tedious and timeconsuming work if feasible at all.

For a broad application study,a flexible software is needed.In the followingsectionstherearethreewaysdescribing,creatingandtraining image analysis algorithms on previously captured images,as they were used for the analysis monitoring in the presented studies:manual;automated circular analysis(ACA)as well as automated complex structure analysis.Measured diameter units are generally converted from pixels to micrometer.These absolute values can be used to calculate the transient Sauter Mean Diameters of droplets as well as the drop size distribution of the system.d32is generally calculated like:Σ d3i/Σd2i,following ISO-standards,see for example ISO 13322-1:2014,the specification diameter is only valid for spherical particles and should be used as a length x instead for particles in general,and therefore d32=x1,2according to ISO 13322-1:2014.

The area weighted mean size x1.2,also known as Sauter mean d32,is approximately

The arithmetic mean size x1,0,for spherical particles often displayed as d10,is approximately

Percentiles are typically used to characterize particle distributions.To determine the percentile value of the p-percentile with 0＜p≤100 the method of the Nearest Rank is applied.Number based percentiles are indexed with n,indicating the number,and a percent value X.The percentage X represents the percentage of particles,whose sizes are less than xnX.For example:a characteristic diameter dn10=15 μm indicates that 10%of all particle diameters are less than 15 μm and 90%are greater than or equal 15 μm.Volume based percentiles are indexed with a small letter v,indicating the total volume of the population,and a percent value X.

Despite different mean diameters and percentiles,calculated from the population of particles of one data point,also the full distributions are analyzed.The non-negative density distribution function qr(x)represents the likeliness of the occurrence of a specific event x,e.g.,a certain particle size.The cumulative distribution function Qr(x)is the integral of the density distribution over the interval[xmin,xi].The index r reflects the typeof quantity of a distribution used for calculation(see Table 1).Themathematical relationship betweenQr(x)and qr(x)is

By definition,Qr(x=xmin)=0 and Qr(x=xmax)=1.

3.2.1.Manual analysis

The manual evaluation needs to be done by four different testing persons at the same time to avoid the influence of personal subjectivity[68].Since the evaluation is based on statistical methodology,the number of dispersed particles must be large enough to generate a valid size distribution[69].For standard systems with spherical particlesthisnumberisusuallya coupleofhundreds[3,11,70],however for complex multiphase system with broad distribution in size,shape and morphology,a statistical reliably number of investigated particles can be a couple of thousands.One measurement point is often characterized by the recording of about more than 100 images.

3.2.2.Automated circular analysis

The automated evaluation of the particle sizes is carried out with the SOPAT software[Version 2.1.11].Manually acquired particles(three to a maximum of twenty)have to be provided to the program,as the benchmarks for the following step of automated search by the software.The search pattern generated from the sample particles marked by the user is compared with the pre-filtered images in the image analysis steps.Thisisfollowedbyacalculationofpotentialcentersoftheparticlecandidates andthena reviewof them byanexact boundaryinvestigation.The user can significantly influence this search process since the generated pattern depends significantly on the choice of the marked particles.Most suitable for the creation of the pattern was the marking of less,but representative particles with uniform grey value over the edge.As a result,a high hit rate can be achieved with very few error detections.As an output of the automatic SOPAT software,the user also receives a tabulated summary of the detected radii and positions of the particles on the image in the form of a CSV(comma separated value)file as well as a graphical output of the recognition.Example images with recognition results of the automated circular analysis(ACA)in a highly concentrated liquid-liquid system can be found in the following Fig.6.Before checking a grayscale arrangement in the image for pattern matching[71],variousimageprocessingalgorithmsareappliedtotheimageseries.The noise in the pictures is reduced by the self-quotient image method[72].For more technical details see Maa? et al.[68].

3.2.3.Automated analysis of complex image structures and complex particle geometries

A number of image analysis approaches also use segmentation algorithms for irregular shape particle analysis(IPA),that are not limited to spherical particles and enable the user to obtain shape and morphology descriptors.Opposed to the spherical analysis,which is based on pattern matching,this analysis is based on segmentation and classification.Segmentation is used to identify structures in an image.It is unconstrained by search for a regular shape,which allows structures with complex boundaries to be detected.Panckow et al.did investigations of irregular objects using image analysis by increasing the shape complexity starting from spheres[27].An improved version of the described IPA is now developed by SOPAT(see results in Sections 4.5 and 4.6)[73].

After the initial image pre-processing,segmentation is carried out to identify the structures in the images(segmentation).To decrease the noise,single pixels which are surrounded by white pixels are deleted,while also the particles as small as one pixel are neglected.From the identified segments 2D information,for example area and the full pixel by pixel boundary information,is captured for a precise shape analysis.The identified particles then need to be classified(for example crystals,bubbles,droplets,sharp and blurred,undefined objects).The number of different classifications can be chosen and is usually three to five.For each of the particle class a feature vector is created.The feature vector(i.e.sphericity,min.Intensity,convexity)is then compared and adjusted to the labeled target particles.After this training procedure segmented particles are compared and only particles with the accepted feature values are retained(Classification).The results of the analysis are saved in a similar way to the results from the spherical analysis algorithm.

Fig.6.Example images with recognition results from a highly concentrated liquid-liquid system.

The quantification of irregular shaped particles is usually carried out using the Feret diameter.Therefore,the characteristic values are given for three different diameters:the minimum Feret,the maximum Feret,and the mean Feret diameter.The Feret diameter(also called“caliper diameter”)is based on the distance between two parallel planes that restrict the particle or,rather,the particle projection(see Fig.7—left).The minimum Feret diameter dF,minis the smallest Feret diameter obtained by rotating the particle gradually to 180°between the parallel lines.

As the Feret diameter is the maximum length of the particle projection on a measuring plane represented by the vector.For a given polygon with N cornersand a normalized measuring plane vector(i.e.=1)the Feret diameter is the difference between the minimum and maximum projection upon the normalized vector

Here·isthedotproduct.TocalculatetheFeretdiametersdF,min,dF,maxand dF,meanthe measuring plane vectoris rotated equidistantly in 16 steps by default from 0°through 180°for each particle.

In the case of spherical particles,of course the maximum is equal to the minimum Feret diameter.The differences for irregular particles between the minimum and maximum Feret diameters are displayed with real crystal examples in Fig.7(right).All particle characteristic connected calculations in the SOPAT software are based on the ISO standards ISO 9276-1:1998 and ISO 9276-6:2008.

3.2.4.Automated analysis of different kinds of particles in parallel

Dispersesystemsoftenconsistofmorethanonlyonedispersephase(see also Fig.8 for examples),sometimes by definition of the process,sometimes unwanted such as entrapped air bubbles in an emulsification process.These unwanted air bubbles do not necessarily influence the product properties nor its quality,however they do influence the measurement signal of all market available technologies significantly.

The automated images analysis is the valuable method to overcome this challenge.Different substances,different phases will appear differently on the image based on the difference in their optical properties.Gas-liquid interfaces usually have high values for the refractive index and will therefore appear with strong contrast.Fig.8 gives such an example image with the connected analysis results as well as with the used image analysis workflow(IAWF).The shown image(Fig.8—left)was recorded in a stirred extraction column.The disperse phase is toluene and the continuous phase is water.Due to some challenges in the process,air is introduced regularly into the process(darker particles in Fig.8).These entrapped air bubbles do lead to a significant falsification of the measurement signal and thereby to a false interpretation of the mass transfer in the column.Therefore,the SOPAT system is introducing a differentiation between the droplets and the bubbles.The different image analysis steps are described in the flow chart(Fig.8—right).

I.First step:all images for the data point are recorded

II.Second step:the images are analyzed three times:

1.All circular objects are recorded(toluene droplets and air bubbles together)

2.All dark circular objects are recorded(air bubbles)

3.All ellipsoidal objects are recorded(toluene droplets)

III.Thirdstep:thedark objects(airbubbles)are subtracted from the population of particles of interested

IV.Fourth step:the ellipsoidal and circular objects(now only toluene droplets)are combined as the full population of all particles of interest(Fig.8—right).

Fig.7.Explanation of different diameter types(left)and the differences between minimum and maximum Feret diameters(right)are displayed.

Toensurehighqualityresults,thesoftwareisonlyanalyzinginfocus objects with sharp edges to avoid size under or over representations by out of focus objects.

3.3.Experimental overview

As already discussed in Section 2 and shown in Table 2,the application of multiphase systems with the need for inline measurements is very broad.Therefore,various applications with increasing complexity have been targeted to demonstrate the reliability as well as the preciseness of the image analysis(IA).Different experimental set-ups have been chosen.All experiments are summarized in Table 2.

Fig.8.Example image analysisresult,using a workflow(WF)with multiple image analysissteps,differentiating bubblesand droplets(left)explanation of the used workflow inform of aflow chart(right).

The first experiment was focusing on liquid-liquid studies and conducted in a mixer-settler unit with kerosene as dispersed phase and water as continuous phase.The stirrer speed in the mixer as well as the dispersed phase fraction ? were varied.The dispersed phase fraction,also known as holdup is determined as follows:

As the experiment is operated continuously,the holdup is influenced by the flow rate FR:

where Fcis the volume flow rate of the continuous phase and Fdis the volume flow rate of the dispersed phase.

The second set of experiments was also focusing on liquid-liquid application,but in a continuous reactive system where mixing was performedbyastaticmixer.Duetotheunwantedentrainmentofairbubbles,they were subtracted using a standard subtraction WF.The third experiment displayed in the following section is again a reactive system.The continuousphaseiswaterandthedispersedphaseisstyrenewhichpolymerizesintopolystyreneduringthereaction.Again,polymerparticlesare analyzed in the fourth set of experiments.Not only do the size of the particles and their distribution need to be monitored,but also their concentration.Therefore,the calibration of relative to absolute concentrations(CoCa,concentration calibration)is discussed and the results are presented including example images,representing the clear visibility of this operation parameter on the displayed image details.Another reactive systemisinvestigatedinexperimentsetno.five.Theconcentrationofcatalyst particles and their size distribution are analyzed.The continuous phase is liquid wax at temperatures between 135 and 180°C.The most complex IA is carried out for the set of experiments number six(Table 2—e06).The liquid phase is water in an air lift loop reactor[74].The whole process is a crystallization process.The air in the loop reactor is used to gently mix the system without mechanical stress on the crystals.The crystal growth has to be tracked separately from the bubbles as a certain size range of the crystals is needed for a successful process.In this study three classifications were made(1—clear crystals;2—air bubbles;3—blurredorunclearobjects)bylabelingupto20 exampleobjects manually.The IAWF is shown in Section 4.6 as well as the bubble size,the crystal size distribution and its shape distributions.

4.Recent Extensive Applications

In the following,the potential of inline photo-optical measurements for various applications is summarized.Each experiment has been conducted in a specific setup(see Table 2).The photo-optical instrument had to be adjusted in regard to magnification,focus position and light intensitytomeettospecificrequirementsforthedatableparticlesizerange.Theadaptabilityofthesystemsmakesitpossibletodeploytheminawide rangeofapplicationsprovidingtargetedprocessinformation(PSD,shape,concentration)of specific particles in its process environment.

4.1.Experiments in a mixer-settler(e01)

Experimentsinmixer-settlerwithKerosene(ρd=820.1kg·m?3,μ=0.00204 Pa·s)as dispersed phase and deionized water(ρc=993.8 kg·m?3,μ =0.00103 Pa·s)as continuous phase are shown.The measured interfacial tension is σ =0.0041 N·m?1.As no surfactants were used,the system does coalesce very quickly.The settling times for the whole set-up after agitating with the highest stirrer speed are less than 3 min.All images were quite similar in appearance.One can see the particles homogeneously dispersed in the liquid shown in Fig.9.Most of the detections are successful.The images are organized as double images:on the bottom the original images after filtering and contrast enhancementareshown,onthe top the original images including detection results are displayed.

The quantitative results of the influence of the stirrer speed as well as flow rate and dispersed phase fraction for the introduced kerosene/water experiment are displayed in Fig.10.

Fig.9.Two example images which show the detection success of the automated image analysis.The images are organized as double images:on the bottom original images after filtering and contrast enhancement are shown,on the top the original image including detected(encircled)particles.

Fig.10.Influence ofthe dispersedphase fraction(left)and thestirrer speedontheSautermean diameter(right).The increaseof dispersedphase fraction for a constant stirrer speeddoes lead to an increase in drop size due to higher coalescence,the increase of stirrer speed at a constant flow rate(FR)does lead to decreasing drop sizes due to increasing energy input.

Fig.11.Image processing and object recognition steps,1—original image,2—enhanced image,3—object recognition,4—original image,5—object recognition with a workflow to subtract all occurring gas bubbles(here red circled).

TheSautermeandiameterisdependingonthedispersedphasefraction ?.The higher amount of kerosene does lead to higher coalescence probabilities and to larger coalescence rates[75].Additionally,theincrease in dispersed phase fraction increases the dampening of the turbulence in the system[76],leading to lower breakage rates.Both phenomena are responsible for the clear increase in the Sauter mean diameter with increasing holdup.

Table 3 Particle characteristics of the experiment 2(e02)comparing the overall PSD(dF,with bubbles)with the drop size distribution(dF,without bubbles)

All four different stirrer speeds show clearly a linear behavior between the Sauter mean diameter and the dispersed phase fraction,which is in good alignment with literature[75].

The results for the variation of the flow rate are shown in Fig.10—right.Thedecreaseoftheflowrate is a decrease inthekerosenecontent in the system(see Eq.(6)).Therefore,the decrease in FR must lead to a decrease in the drop size according to the above discussed phenomena.This can be seen clearly in the displayed diagram.d32for a stirrer speed of500r·min?1decreasesfrom573μmataFRof5/1downto323μmata FRof 1/5(inversed flow rate).The influence of the stirrer speed on d32follows the expected behavior.The increase in n does lead to a decrease in d32.The relation follows the expected relation d32～n?C[75].These reliable results do show the reliability of the inline measurements of fast coalescing droplet using a photo-optical device.

Fig.12.Changing image quality and objects detection(green circles)in the Polymerization,start(left)-after initiation(middle)-end(right).

4.2.Discrimination of gas bubbles in a liquid/liquid/gas dispersion(e02)

This experiment was performed directly after a static mixer before the reaction chamber.The concentration of the dispersed phase was varying between 10 vo1%and 30 vol%during the experiment.It was known that entrapped gas bubbles influence the drop size distribution,beingunabletodiscriminatethegasbubblesfromdropsizedistribution with anygivenparticle analyzer.Hence theinfluence of the static mixer on thedrop sizedistributionis biased.This study shows theinfluence of gas bubbles and importance of a targeted particle detection.With a trained WF(see Section 3.2.4)excluding the gas bubbles from the particle size distribution(PSD)it was possible to compare the overall PSD with the drop size distribution of the 1st disperse phase(Fig.11).

In Table 3 the effects on particle size distributions are shown when gas bubbles are excluded from the overall distribution in comparison to the overall distribution.Especially xv90shows a significant deviation to the between the two distributions.Having a more distinguished knowledge of the dispersed system the flow rate can be reduced to achieve the required xv50.

4.3.Monitoring of a polymerization process(e03)

In this study a suspension polymerization was monitored over a timeof 2 h in a labscale temperaturecontrolled DN100 baffled glass reactor.Thetotalvolumeoftheset-upis500ml.Thedispersedphasefraction is 48%of the volume.The chemical composition of the dispersed and aqueous phases as well as the initiator is under confidentiality.The viscosity of the dispersed phase is around 0.77 Pa·s,the viscosity of the aqueous phase is close to water viscosity.The system is mixed throughthewholeexperimentwiththesamestirrerspeedandconstant temperature of 70°C.After 75 min the initiator is added into the aqueousphase.Thequantitativeresultsandtheinfluenceoftheinitiator for this experiment are displayed in Fig.12.

The Sauter mean diameter as well as the arithmetic mean size increase rapidly at around 10%after initiation(survey point 27)and start stabilizing already 10 min later after survey point 31(Fig.13).

4.4.Concentration correlation of polymer beads(e04)

At the end of a typical bead suspension polymerization process for expanded polystyrene aside from the particle size distribution of thefinal product,the concentration of the solid particles in suspension is a crucial parameter.In this experiment a photo-optical device was inserted into six different concentrations of a suspension with the same type of micro-beads(seeFig.14).Fourexperiments withdifferent probe configurations are shown in Table 4.

For each concentration,four image sequences of 100 images have beenacquired.Duetotheillumination and optical properties of thesystem refraction increases with higher concentration,whereastheoverall brightness and sharpness as features of the acquired images decrease.Knowing that the LAPD(arithmetic mean value of the Laplacian operator)can be used as an indicator for the sharpness and the standard intensity an indicator for the brightness of an image,for every image sequencefirsttheLAPD1https://de.mathworks.com/help/matlab/ref/std.html;access on 29.11.2017 at 6:33 pm.,2https://de.mathworks.com/help/images/ref/imfilter.html?searchHighlight=imfilter&s_tid=doc_srchtitle;access on 29.11.2017 at 6:34 pm.isgeneratedbytheconvolutionoftheLaplacefilter followed by arithmetic averaging.The calculated elements are subsequently divided through the standard deviation of the image intensity3https://de.mathworks.com/help/matlab/ref/std.html;access on 29.11.2017 at 6:35 pm.(StrdInt).The results of this calculation for each image sequence are considered as the relative concentration.In all experiments therelative concentration showed a linear correlation with a coefficient ofdeterminationofhigherthan0.96totherealconcentration.Havingat least three real particle concentrations of the suspension,allows this method to interpolate and predict all concentrations with a standard deviation s2of max.1.55.

This study has shown that photo-optical instruments operating in transmission illumination method(see Fig.5)can be calibrated to automatically compute the particle concentration of solid particles based on a feature analysis(see Fig.15).

4.5.Catalyst particles in liquid wax(e05)

The concentration was not calculated based on a feature vector for this experiment but on the actual number and size of the detected particlesusingthepatternmatchingalgorithm.Asalwaysthisisonlyalocal measurement which gives the concentration at a certain time for a certain position in the reactor.Therefore,the system could also be used to studythespatialdistributionofthecatalystconcentrationinthestudied vessel.

Fig.13.Behavior of arithmetic mean size x1,0and Sauter mean diameter in a suspension polymerization over time.

Fig.14.Example image of every particle concentration from Experiment 1.

The concentration is calculated as follows:Every sample is analyzed based on50imageswhichwererecordedwithaframerateof 10 images per second.Those 50 images are analyzed using the pattern matchingalgorithm[68].Now the absolute cumulative number of all particles on all images can be compared with each other.The experiment was carried out twice to prove the reproducibility of the data,and the standard deviation of the two measurements is only 1%.The values for the averaged particle number and its standard deviation are given in Fig.16.

Table 4 Hardware configurations of the photo-optical device

Although the concentration is constantly increasing from sample A to sample F,that was not reflected in the particle number.Sample A(with an average of 377 particles)to sample E(with an average of 26508 particles in one set of images)do show a significant increase in particle number.These data are clear and sound.Sample F shows anaverage particle number per data point of 26323.

Fig.15.Concentration correlation in Experiment 1 of the six different concentrations.

Agglomeration can bind critical catalyst particle mass fraction in larger particles,therefore the real concentration has to be calculated from the number and size of particles.The mass concentration can be calculated from the sum of the individual particle volumes times their density related to a control volume.Since the density of the catalyst particles and the control volume can be assumed to be identical for all experiments,acomparisonoftheconcentrationsislimitedtoacomparison of the cumulative sum of the individual particle volumes.If the particles are considered as spheres for simplicity,the volume can be replaced with a representative diameter.

The results in Fig.17 show the comparison of the products of number of measured particles times a representative diameter(d10,dn,50,dn,90,dv,50).The product shows a clear trend.In particular,the differentiation between sample E and sample F is now much clearer than when considering the number of particles alone.Including volume and density,as already discussed,a real concentration could also be calculated.For the sake of simplicity,the apparent concentration is used in this study.

Fig.16.Absolute particle number of all six investigated samples.

Fig.17.Catalyst particles inliquidwax(135-180°C,0.6 to1.9 MPa),onechallenge hereistodifferentiatecatalyst particles from trapped air(left);concentration ofcatalyst particles insix different samples(right).

The dn,90gives the clearest tendency—all concentrations can be clearly differentiated from each other,the increase from A to F is clearly visible.Concentration F is only slightly higher than concentration E—the error square of the product with the d10as well as with the dv,50in relation to the original concentration based on weight measurements is the smallest.Theresults with the dn,50led to the worst outcomes.Based on these findings,the measurement of the concentration of one dispersed phase can be easily approximated with the number of particles times a characteristic diameter.This approximation is valid if the particles are close to spherical shape.That was true for the investigated catalyst particles(Fig.17—left).The catalyst particles are homogeneously dispersed and they were clearly detected by the used image analysis.The system is robust to capture sharp images even under high pressure andtemperature conditionsandthereforecanbeused forprocessmonitoring in these kinds of industrial applications.

4.6.Crystallization and discrimination of entrained gas bubbles(e06)

In case of crystallization processes,the various phenomena such as growth,attrition or agglomeration of crystals affects need to be detected.In production of active pharmaceutical ingredients(APIs),it is very much needed to control API properties,such as the shape,size,polymorphism and purity during the crystallization process,also in respectto theQbD(quality by design)philosophy[77].As a standard procedure to influence the crystal size by agitation,often the entrainment of gasbubblescannot beprevented.In this study crystals have beenisolated from gas bubbles to be able to show separate crystal and bubble size distribution during a crystallization process.

A number of images(see Fig.18)show the general procedure of an image analysis workflow to isolate crystals providing a more precise crystal size and shape distribution when the crystallization is taking place.First the image analysis algorithm detects gas bubbles with standard patternrecognitionsteps(see also Section3.2.2).Inthe next intermediate step,boundaries of all objects are segmented,which in the following step will be closed using a multi-variate principal component analysis[78].The next step focuses on refining the detected boundaries andrejectingtheunwantedobjectsbyclassificationmethods.Inthelast step the overall detected boundaries are subtracted by the detected bubbles.

With the workflow for isolation of crystal detection shape informationcan be ascertained(see Fig.19)not only for gaininga better understanding of this crystallization but also for controlling the crystal size and shape development inline in changing fluid dynamics.

4.7.Intracellular fatty acid concentration in microalgae(e07)

The production of polyunsaturated fatty acids for various applications in health,cosmetics and for animal and human nutrition has gained a lot of interest in recent years.While the acid accumulates in lipid droplets under starvation of nitrogen in the heterotrophic algae Crypthecodinium cohnii,cells increase in their size from about 15 μm during active growth to 25 μm.Since cells are shear-sensitive,any method for the detection of a single-cell size distribution can also identify the portion of disrupted cells,which appear smaller then.This is so far conducted with flow cytometry,as established for cell line cultivation.Inline microscopy was successfully applied to quantify the intracellular DHA(docosahexaenoic acid)content based on the averagecellsize obtained from the individualautomated cell recognition.Similar results were achieved with the offline 3-dimensional holographic microscopy[79].The cells were identified correctly with both methods(Fig.19).Although the prediction of the DHA content was more precise with the 3-dimensional holographic microscopy,both methods are suitable to replace sophisticated gas chromatographic analysis.

Recently,the identification of yeast buds through the differentiation of circular and ellipsoidal particle shapes became feasible[80].The single-cell based determination of the budding index allows determining not only the heterogeneity a culture,but also the growth activity.Such a method yields rather the same information than flow cytometry of unstained yeast cells,but it further provides real images of cells,and most important,can be conducted inline without any efforts of sampling(Fig.20).

4.8.Application summary and overview of application demonstrations from literature

The previous sections have demonstrated the capability of inline photo-optical measurements combined with image analysis.In addition,Table 5 gives a literature overview of the research work that has been pursued so far with the examined photo-optical instrument in this study.That should help the interested reader to gain security if his application might be analyzable with the general method of in-line imaging.

5.Conclusion and Outlook

Fig.18.Objectsegmentation fordetecting crystals,1=original image,2=gas bubble detection,3=boundarysegmentation,4=boundary closing,5=classification and subtraction of gas bubbles.

An extensive literature review was carried out to differentiate and understand existing technologies for inline particle analysis.To obtain industrial relevance a focus was made on commercially available technologies.The independent findings from different research studies judge image-based systems as superior over competing technologies such as laser backscattering or spatial filter velocimetry for particles above 1 μm.A certain type of a photo-optical measurement technique was used in the present work to investigate the particle size and shape distributions in different applications.A number of process relevant characteristics can be determined via image analysis such as deformed particles as well as the particle concentration itself.Depending onthefundamentalgoalsofeachexperimentinthisstudytheinlineimaging system has been adapted with regard to the hardware setup as well as the image analysis workflow.After optimization of the image analysis parameters reliable distribution functions can be generated by the photo-optical measurement technique.However,the detectable diameter range of photo-optical devices is somewhat limited.So far for inline photo-optical probes it is not possible to provide a meaningful resolution of particles in the size range from 0.5 μm up to 10 mm with one instrument.Those limitations become obvious considering the volume density distribution q3by the occurrence of an abrupt upper limit instead of a continuous decay.

Fig.19.Aspect ratio density histogram of a continuous crystallization process,total 19064 crystals.

Measurementstakenoveraverylongperiodinconjunctionwiththe recording of an extremely high amount of image data is intended to minimize the fluctuations of the statistical evaluation methodology.At the same time a big quantity of data is recorded,which requires a cautiousdatamanagement.Duetothelargeamountsofdataforparticlediameters and shape,a basis for sensitivity analyses was given.This shows that photo-optical instruments with automated image analysis processing large quantities of particlesare suitable for processmonitoring.Many applications for inline process control,however,will require further reduction in the time required for automated evaluation.

Incontrasttolaser-opticalmeasuringsystems,whichhaveprovedto be unsuitable in particulate systems with a uniform,smooth phase boundary,photo-optical techniques can be a meaningful alternative for determining particle sizes and shapes.

Nomenclature

c concentration,%

C constant

cequiequivalentconcentration,basedonnumbertimesdiameterof particles,μm

d10arithmetic mean diameter of spherical particles,μm

d32Sauter mean diameter,μm

dF,maxFeret maximum diameter,μm

dF,meanFeret mean diameter,μm

dF,minFeret minimum diameter,μm

dn,5050%-percentile value of number based size,μm

dn,9090%-percentile value of number based size,μm

dv,5090%-percentile value of volume based size,μm

Fcvolume flow rate of continuous phase,m3·h?1

Fdvolume flow rate of dispersed phase,m3·h?1

FRflow rate,

n stirrer speed,min?1

N number of particles

qr(x) density distribution,μm?1

Qr(x) cumulative distribution

ΔQr(x) relative distribution

Vcvolume of continuous phase,ml

Vdvolume of disperse phase,ml

Table 5 Literature overview of investigated biological and chemical systems using the photo-optical SOPAT system

x length,μm

x1,0arithmetic mean size,μm

x1,2area weighted mean size,also known as Sauter mean,μm

σ interfacial tension,N·m?1

ε power input,W

λ wavelength,nm

μ viscosity,Pa·s

ρcdensity(continuous phase),kg·m?3

ρddensity(dispersed phase),kg·m?3

φ dispersed phase fraction

Acknowledgments

The support and the data by Prof.Hermann Kramer(TU-Delft)for the air lift loop reactor,the technical details by Hannu Eloranta(Pixact)and the work in the field of image analysis by Dr.Michael Muthig(SOPAT)are highly appreciated.The development of a new inline microscope has been financially supported by the grants for the project“Smart Process Inspection”(funding code ZF4184501CR5)from the “Zentrales Innovationsprogramm Mittelstand”(ZIM).