PTDLR UF GHK

1
Sounding the Development Potential of Data
Analysis Methodologyfor Metal DectorsJoint
Project HuMin/MD(funded by BMBF)

• Gerd-Henning KleinPT-DLR

2
HuMin/MDOutline of Presentation

• Mission/Objectives
• Reduction of MD false alarm rates of
  off-the-shelf MD
• Methodology of data analysis in subsurface
  sensing
• Tasks/Approaches
• Application of advanced methods of signal
  analysis
• Local 3D Electromagnetic Induction Tomography
• Forward modelling (computational electrodynamics)
• Model-based data inversion (advanced mathematics)
• Consortium
• Conclusions

3
HuMin/MDMission/Objectives

• For the sake of MD false alarm reduction,
  sounding of the development potential of
• A. Data analysis methodology for off-the-shelf,
  hand-held MD
• B. MD sensor technology (building on the results
  of A)
• Reconstruction of features of buried metallic
  objects
• shape, position, orientation 3D EMIT (w/ data
  from local MD scan)
• other (e.g. metal type) signal analysis
  (statistics/informatics)
• Provision/production of necessary measurement
  data
• Virtual data MD process modelling (diverse MD,
  targets and soils)
• Real data MD primary/raw signals (diverse MD,
  targets and soils)
• Application-oriented basic research, with
  practical advice

4
HuMin/MDTasks/Approaches

• Signal analysis advanced methods of statistics
  and/or informatics
• Forward modelling of the MD measurement process,
  using tools of computational electrodynamics
• 3D EMIT model-based data inversion, with
  advanced methods of computer tomography
• Provision of add-on devices for MD data
  acquisition
• for determination of position/orientation
• for sampling of raw signals and
  position/orientation data
• for real-time computation and data visualization

5
Signal AnalysisFeatures Extraction/Classification

• Data preprocessing, e.g.
• Noise reduction (FFT, difference spectra, etc.)
• Transformations (wavelets, etc.)
• Mathematical statistics, e.g.
• Principal components analysis
• Discriminance analysis
• Multi-dimensional scaling, etc.
• Machine learning, e.g.
• Neural networks
• Support vector machines
• Kernel-based learning (Vapnik 1998)

6
HuMin/MDElectromagnetic Induction Tomography

• Data provision
• Local scan around anomaly pin-pointed by MD
  variation of position r??3, orientation W??3 and
  other parameters p??p (e.g. frequency of
  continous wave mode)
• For each configuration ci (ri,Wi,pi)??p6, i
  1,...,m, this scan yields a signal si(t).
• Data analysis
• Forward modelling of the MD measurement process
  (direct problem) with computational
  electrodynamics tools (numerical solution of the
  relevant system of differential/integral
  equations
• Reconstruction of subsurface conductivity
  patterns from scan data (model-based inversion)
  with tomography tools. Resolution depends on
  (usually) irregular pattern of point set c1,...,
  cm.

7
HuMin/MDForward modelling

• Modelling of MD measurement process, building on
  the underlying physics (electromagnetic
  induction, eddy currents)
• Simulations under diverse conditions (MD models,
  MD operations, mines, soils), by numerical
  solution of discretized Maxwell equations (or
  derivates of them)
• Synthetization of virtual MD data, used as inputs
  for model-based inversion and for sensitivity
  analysis
• By-product better understanding of MD
  performance (useful for further development of MD
  technology)

8
HuMin/MD3D EMIT 3 Approaches

• 3 approaches of local tomography followed in
  parallel
• Iteration methods and decomposition
  method(methodologies of Colton, Kress, Potthast
  et al.)
• Linear sampling method and factorization
  method(methodologies of Colton, Kirsch, Hanke et
  al.)
• Equivalent sources and approximate
  inverses(methodologies of Louis, Natterer, Maaß
  et al.)

9
HuMin/MDConsortium

• 10 research/university institutes involved
• Applied/numerical mathematics (tomography, data
  inversion) 4
• Computational electrodynamics (general forward
  modelling) 1
• Exploration geophysics (parametric modelling
  soil database) 2
• Non-destructive testing science/technology (real
  data) 1
• Environmental technology (informatics, signal
  analysis) 1
• Production technology (project management, signal
  analysis) 1
• Practical expertise provided by sub-contractors
• on MD technology (manufacturers)
• on MD use for humanitarian demining

10
HuMin/MDConclusions

• Broad spectrum of rigorous mathematical
  methodologies
• Computational electromagnetics diverse
  discretization approaches
• Tomography novel approaches for model-based
  data-inversion
• Signal analysis novel approaches, e.g.
  kernel-based learning
• Ambitious goals, with several developmental
  risks, but
• Proof-of-concept milestones (with stop-show
  criteria)
• 3D EMIT robustness/resolution of imaging (for
  diverse soils)
• Signal analysis robustness of classification
  (for diverse soils)
• Operations computational speed (real time)
  operator interface
• Experiences potentially useful for RTD in MD
  technology
• A template for analysis of emerging sensor
  technologies?