Michael Schmuker

School of Engineering and Informatics

University of Sussex, Brighton, UK

http://biomachinelearning.net

European Robotics Forum 2016

Cankarjev Dom, Ljubljana

23/3/2016

smoke plume image

by william warby

CC-BY 2.0

INFERRINGSOURCE DISTANCE FROMTHE FINE STRUCTURE OF GAS PLUMESUSINGMETAL-OXIDESENSORS

2

Why estimate the distance of a gas source?

¡ Inbiology:

- Locatingfoodsources,matingpartners,predators

¡ Intechnology:

- Gas-basednavigation, e.g.arobot locatingthesourceofagasleak

- Environmental monitoring, e.g.pollution, wildfires,etc.

The AASS Mobile Robotics & Olfaction Lab in 2015* o 15 Ph.D. students, 12 senior researchers

o 7 ongoing projects, funding > 1.3M€/y

Research Area 1: Mobile Transport Robots for Industrial Applications o 3D perception, robot vision and navigation

o autonomous and safe long-term operation in the real world

o extensive technology transfer through collaborative projects with industrial partners

o autonomous forklifts, wheel loaders, mining vehicles, service robots on airports, in hospitals and for cleaning

o leader in Europe, key player world-wide

Prof. Achim Lilienthal

contact: www.aass.oru.se/~lilien achim.lilienthal@oru.se

Research Area 2: Mobile Robot Olfaction o gas sensing with sensor systems

in open sampling configuration (from electronic to mobile nose)

o gas sensor networks for air pollution monitoring, mobile robots for surveillance of landfill sites, gas leak localization, gas-sensitive flying robots

o world-leading in mobile robot olfaction

*16 Jun 2015 beespotter.org/topics/genome/

A. Lilienthal, Örebro University

www.aass.oru.se/Research/Learning/amll.html

odotech.com

istockphoto.com

3

Gas concentration vs. distance

¡ Naïveassumption:gasconcentrationindicatesdistance.

¡Motivatedbytheconceptofdiffusion.

- Gasconcentration ishighestatthesource.

- Diffusion dilutesthegaswithincreasingdistance.

¡ Problemswiththenaïveassumption:

- Concentrationatdistanced dependsonconcentrationatthesource.Whatifthesourceconcentration isunknown?

- Onlyvalidinabsenceofairmovement (i.e.,wind)andturbulence.Notrealistic.

Image source: Ursell, T. S. (2013). The Diffusion Equation: A Multi-

dimensional Tutorial.

4

The structure of gas plumes

¡ Gasplumeshaveacomplexstructurethatcanbecharacterised onthreelevels:

¡ Time-averaged:downwindparabola-likeshape.

¡ Temporalmesoscale:Plumemeandering

¡ Fine-scale:filamentous.

Ann

u. R

ev. E

ntom

ol. 1

992.

37:5

05-5

32. D

ownl

oade

d fr

om w

ww

.ann

ualre

view

s.org

by W

IB60

13 -

Frei

e U

nive

rsita

et B

erlin

- FU

Ber

lin o

n 10

/13/

11. F

or p

erso

nal u

se o

nly.

Murlis, J., Elkington, J. S., and Cardé, R. T. (1992). Annu. Rev. Entomol. 37, 505–532.

5

Intermittency vs. distance in a wind tunnel

¡Windtunnelexperiments:

¡Mostsalientdifferenceisthechangeinintermittency.

¡Max.concentrationisalsoaffected.

¡ Averageconcentrationwouldbeinfluencedbybothfactors.

128 KRISTINE A. JUSTUS ET AL.

Figure 7. Representative traces from the axial center of the continuous plume using a minia-ture photoionization detector (miniPID) at four distances downstream of the odor source: (a)50 mm, (b) 100 mm, (c) 200 mm, and (d) 400 mm. Each trace is 15 s.

between mean concentrations at 200 mm and 400 mm. The concentrations at 200and 400 mm are approximately 20% of the concentration 50 mm from source.

The standard deviation of the fluctuating component of the concentration signalis a measure of the ‘intensity’ of the concentration fluctuations and is shown hereas a ratio to the local mean concentration. We have estimated this ratio for each ‘x’position by plotting the standard deviation of the fluctuating concentration againstthe mean taken from each of the records for each of the (12 by 12 mm) cross sec-tions and calculating the slope of a linearized approximation to the relationships.This enabled us to combine data from across the whole core regions of the plumes.

Justus, K. A., Murlis, J., Jones, C., and Cardé, R. T. (2002).

Environ. Fluid Mech. 2, 115–142.

d = 50 mm

d = 100 mm

d = 200 mm

d = 400 mm

PLUME STRUCTURE IN A WIND TUNNEL 119

Figure 2. A 3-m long wind tunnel typical of entomological studies, constructed withLexan! and Plexiglas!. The upwind end has a 150-mm deep block of aluminum Hexcel! tolaminize airflow. Window screening covers both the upwind (not shown) and downwind endsof the tunnel. A cowling of polyvinyl sheeting attaches the downwind end to an exhaust duct(not shown). Openings in the side of the tunnel are sealed with polyvinyl sheeting to minimizethe disturbance in airflow during experiments.

meaningful conclusions about changes in plume characteristics over a region ofplume that, in terms of distances in the field, is close to the source. Our eventualaim is to understand how moment-to-moment contact with filaments of pheromoneinfluences the moth’s flight maneuvers.

The system we have developed is not optimized for the study of fluid mechanicsof dispersing plumes but rather is a practical system to determine fine-scale featuresof airborne odor plumes that have been commonly used in insect investigations forseveral decades. In such highly complex arrangements optimized for the study ofinsect flight, there remains no substitute for an empirical approach in characterizingthe stimulus that insects receive.

2. Materials and Methods

2.1. WIND TUNNEL

The wind tunnel used in these trials was a typical insect flight tunnel similar tothat used by Mafra-Neto and Cardé [1, 14, 15] and Cardé and Knols [21]. It wasconstructed from three sheets of 3 mm thick clear polyacrylic Lexan!. To make thetunnel, each sheet was bent into an arch-shape and connected to a 1-m-wide frameto create a symmetrical semi-cylinder 3-m in length with a center height of 1.5 m(Figure 2). The floor was constructed from 10 mm thick clear Plexiglas! sheets. Apush-pull airflow system was used with variable speed motors operating an upwindpusher fan and downwind exhaust fan.

Turbulence in the tunnel, as indicated by flow visualization, was greatly reducedby drawing air into the tunnel through a 150-mm-thick aluminum honeycomb,(Hexcel®, 15 mm diameter cells), 200 mm downstream of the upstream fan. Twooblong openings were cut into one side of the tunnel, 270 mm from the upstream

6

Exploring the physical component of odour space

¡ Vergaraetal2013:Electronicgassensorsinaturbulentwindtunnel.

- FigaroTGS26XXmetal-oxidesensors

¡ Sensorsatvaryingdistancefromtheodour source.

¡ Publicdataset,18k72-dimensionaltime-seriesrecordings, approx.120GB.

A. Vergara et al. / Sensors and Actuators B 185 (2013) 462– 477 465

Fig. 1. The custom-designed block diagram used as the chemical detection platform, containing nine portable chemosensory modules (top). Front and rear view of thecustom-designed portable metal oxide based chemical sensory module (bottom, left and right, respectively). Each chemosensory module is endowed with eight metal oxidebased gas sensors and the necessary signal conditioning electronics and can be placed in different locations of the wind tunnel. In the board, the operating temperature ofthe MOX sensors is controlled by means of a PWM whereas the sensor resistance response is indirectly measured from a standard measuring circuit for an optimal signalconditioning (10 k! load resistor) before the signals are acquired with a 12-bit resolution ADC.

Fig. 2. Wind tunnel test bed facility used to collect time series data from sensor arrays for the problem of gas identification. The red point in the schema indicates the locationof the chemical source. The displacements of the six training lines (or positions) are labeled in the schema as P1–P6. Each training line includes nine landmarks evenlydistributed along the line to complete a grid of 54 evaluation landmarks. Notice also that the smoke distribution shown in the wind tunnel is a dramatization that illustratesthe dispersion mechanisms of the gas emissions within the wind tunnel. An actual distribution map of the chemical analytes being released is graphically illustrated in Fig. 5.Finally, to measure the ambient temperature and humidity during the entire experiment in the wind tunnel we utilized a sensor tagged by the manufacturer, Sensirion, asSHT15.

7

Turbulent gas flow

¡ Concentrationdecreaseswithdistancefromsource.

¡ Butinordertopredictdistance,theconcentrationatthesourcemustbeknown.

¡ Inaturbulentenvironment,gasintermittencyalsodependsonsourcedistance.

¡ Canweexploitsuchspatio-temporalfeaturesofgasplumesfordistanceestimation?

8

Bandpass to obtain the “interesting” part of the signal

¡ Criticalfrequencies0.025Hz,0.5Hz.

¡ Lowband:gasflowon/off.

¡ Highband:noise(mostly).

¡Middleband:?

9

Spectral analysis

¡Middlebandvariestheslowerthefartherfromthesource.

¡ Calculatingrelativepowerinthefourier spectrum:fcrit =0.1Hz

¡ Relativespectralcontentofhighfrequencypredictsdistancefromsource.

Schmuker et al., submitted.

10

Bout-based distance prediction

¡ 3-stagefiltering: low-pass->differential ->leakyintegrationw/exponential.

¡ Reveals“bouts”:portions ofthesignalwithpositiveslope.

¡ Boutcountspredictsourcedistance.

Schmuker et al., submitted.

11

Cross-wind statistics

¡ Variance of bout count indicates

crosswind offset from plume centreline.

12

Potential strategy for gas-based robot navigation

13

14

Students, Collaborators & Funding

Students/Postdocs:JanSölter PhD,FUBerlin(nowsoftwaredev,Berlin)

TaraDezhdar cand.PhD,FUBerlinIuliaLungu cand.MScBCCNBerlin

ViktorBahr BachelorBioinf,nowTUBerlinBahadir Kasap MSc,nowcand.PhDDonders Inst.,Nijmegen

BenjaminAuffarth postdoc,nowFORTH,Heraklion

StephanGabler MScBCCN,nowsoftwaredev,SFBayArea

MarcusSchroederMScBioinf,softwaredev,Berlin

SPP 1392: Integrative

Analysis of Olfaction

Collaborators:Gassensors&bioinspired machinelearningAlanDiamond,ThomasNowotnyUniversity ofSussex, Brighton,UKRamonHuertaBioCircuits Institute,UCSanDiego,USA

MouseolfactorybulbJanSchumacher,Hartwig SporsMPIforBiophysics, Frankfurt(nowUniversityHospitalGiessen,Germany)

NeuromorphicHardwareThomasPfeil,KarlheinzMeierKirchhoff-Institute forPhysics,Uni Heidelberg,Germany

InsectolfactoryphysiologySilke SachseMPIforChemicalEcology,Jena,Germany

NeuronalcodingofpainGaryLewin,Rabih MoshourabMax-Delbrück-CentreforMolecularMedicine,Berlin

Funding:

Marie Curie IEF