corpral development ivan galkin, bodo reinisch, grigori khmyrov, alexander kozlov, xueqin huang,...
TRANSCRIPT
CORPRAL Development
Ivan Galkin, Bodo Reinisch, Grigori Khmyrov, Alexander Kozlov, Xueqin Huang,
Robert Benson, Shing Fung
Outline
• Corporal development– Grant from NASA Intelligent Systems (IS)
Program– New equipment
• Cerebral development– New resonance matching algorithm– Cooperation with U.Penn Institute of
Neurological Sciences
RPI Plasmagrams
NASA IS Program
CICT
cict.nasa.gov
IS
CNIS
SCITSR
AR
HCC
IDU
DM KD
ML
IT strategic research Space comms
Computers, Networks,Databases
Intelligent systems
Automaticreasoning
Human-centeredComputing
Intelligent Data Understanding
IS and RPI data
• Near-term goal: CORPRAL analysis of RPI data for a variety of magnetospheric echoes
• Long-term goal: use of CORPRAL-derived data to modify operating state of onboard instruments
Near-term tasks
• Use IS technology for plasmagram processing– Resonance identification– Trace extraction
• Use expert knowledge to automatically interpret RPI data– Echoes– Resonances– Spectrograms
• Introduce “state of the magnetosphere” index for space weather alerts
• Onboard ML decision making
ML onboard: ideas?
• Magnetospheric State index for Space Weather applications
• Intelligent data reduction• Dynamic antenna tuning
DEMETER
Server Room Setup
DIDB
Athlon MP 2200+
RPI
Pentium-4 900 MHz
CAR
Pentium-4 1GHz
Applicationserver
Duron 700 MHz
CORPRAL
Firebird DBMS(database)
File ServerInstallation WWW pageLatest RPI data page
CVS RepositoryInterclient server for DIDB
RPI LZ ArchiveTOMCAT Server
DIDB IngestionRPI Ingestion
CORPRALADRES
Picture of the day
Firebird DBMS(database)
Interclient Server
ULCAR
Pentium-II 266 MHz
ULCAR HomepageFTP Guest area
DigisondeIncoming
Pentium 200MHz
FTP IncomingDispatcher MP
Digi data archive
5 5 1
Resonance Processing
Resonance Model
1. Gyrofrequency, fce, and its harmonics, nfce
2. Plasma frequency, fpe
3. Upper Hybrid resonance, fT
4. Q-type and its harmonics, fQn (a.k.a. Bernstein mode resonances)
5. D-type and its harmonics, fDn
22cepeT fff
2
2
2
46.0
ce
peceQn f
f
nnff
22
22
95.0
ceDnDn
ceDnDn
cepeDn
fff
fff
nfff
fpe and fce drive all frequencies
Model Fitting Approach
• Superimpose a Comb Template on the plasmagram and seek the best quality of fit for varying drivers fpe and fce
• Used for ISIS, ISS-B, ISEE-1 data• Not good enough for RPI
– Tremendous variety of conditions– Frequency coverage not optimal– Accuracy vs precision issues
• 0.7% accuracy is required for model self-consistency
– Noise environment
Resonance Recognition
Not all peaks of summary amplitude are resonances
Improvements to resonance fit
• Image filter to highlight resonance signatures
• Resonance signature detection and tagging
• Limit contributions to the fit quality to signatures only
Resonance Signature Filter
F(Ai) = median{Aj}, j = (1, i)
(“cumulative” median filter)
0
10
20
30
40
50
60
70
80
90
100
1 5 9
13
17
21
25
29
33
37
41
45
49
Input
Output
0
10
20
30
40
50
60
70
80
90
100
1 5 9
13
17
21
25
29
33
37
41
45
49
Input
Output
FILTER RESPONSE TO A PULSE
Filtering noisy data
0
10
20
30
40
50
60
70
80
90
100Input
Output
0
10
20
30
40
50
60
70
80
90
100Input
Output
Signature Detector
Detection of signatures allows evaluation of their contrast that is then used to calculate quality fit instead of amplitude
Change of medium during sounding
• The “driving” fce and fpe are specified at the plasmagram start
• Gradients of fce and fpe are estimated using the model values at start and stop times
• An iterative scheme is applied to ensure that templates are placed at the frequency that is compatible with fce and fpe at that time
Gradients of fp and fe
fce(t1) fpe (t2) fT (t3)
)()()(
)()()(
32
32
3
12
22
3
tftftf
tftftf
cepeT
cepeT
1. Driving fce and fpe are taken at the plasmagram start2. Template frequencies are corrected for the gradients of
the driving fce and fpe
3. Iterative scheme is used to find self-consistent set of all involved resonance frequencies
Time, frequency
Resonances in BB
Marr’s Paradigm
Raw Image
Echoes
Traces
Classified Traces
Decisions
RotorsSaliency Map
From raw image to echoes
RAW IMAGE LABELING (NO THRESHOLDING)
LABELING (AFTER THRESHOLDING)
ADAPTIVE THRESHOLDING (a.k.a. ECHO DETECTION)
Rotors
• Rotors – local estimates of line orientation at each labeled pixel
• Orientation estimates are subject to errors (due to the range jitter)
Saliency
• Saliency measure:– How likely the rotor belongs to a
contour
• Saliency map– Image, where each pixel intensity is
its saliency measure
Gestalt principles
Key principle for contour saliency is continuity
Us and Them
• Us:– Rotor orientations
are estimated using directional histogramming
– Saliency map is obtained by iterative optimization in the network of rotors
• Them:– Rotor orientations are
obtained using steerable filters (e.g., Gabor filter banks)
– Saliency map is obtained by cumulative contribution in the “cortical” network of rotors
Cortical networks
Facilitation term for rotorsin the model of striate cortex
[Yen, Finkel, 1997]
Rotor Optimization
End of OptimizationStart of Optimization
Hopfield NN optimizer
Structure of neuronNobel Prize [1906]
Multilayered Perceptron(feed forward, back-propagation NN)
Feed-back Hopfield NN
Inp
ut
Ou
tpu
t
Rotor Optimization
CO-CIRCULARITY CONSTRAINTa.k.a. Prägnanz, principle of curvature constancy in Gestalt
Striate cortexmodel
[Baginyan et al., 1994]
Near Zone
Range jitter deteriorates facilitation from nearby rotors
Parasitic stable state
“Tunneling” through energy barriers using MFT approach(introduction of thermodynamic noise in the NN evolving rule)
Perceptual Grouping
• Us:– Bottom-up
clusterization using rotor interaction as the distance criterion
• Them:– Synchronization
(“together”) and desynchronization (“apart”) in a cortical network