image processing in magnetic fusion devices · 2019-06-23 · image processing in fusion 04.06.2019...
TRANSCRIPT
04.06.2019 1
Marcin Jakubowski
Max-Planck-Institut für Plasmaphysik, Greifswald, Germany
Image Processing in Magnetic Fusion Devices
Images in science
04.06.2019 IAEA Meeting on Fusion Data Processing, Validation and Analysis, 30th May 2019 2
• Visual information has always played an important role during the evolution of science.
• The advent of the imaging in science came with digital cameras, which allowed for easier, faster and more accurate detection of events in magnetic fusion devices.
C. Darwin, “Voyage of the Beagle”
C. D. Anderson, Phys. Rev. 43 (1933) 491
TFTR
Disruption caused by a locked mode at TFTR. Intensified image in the wavelength range of 440–700 nm, 2000 fps/30usec exposure, Each full image is composed of 239 ×192 pixels and the digitalization is 8 bit deep
R.J. Maqueda and G.A. Wurden, Nucl. Fusion, 39 (1999) 629
Challenging definition of an image
04.06.2019 IAEA Meeting on Fusion Data Processing, Validation and Analysis, 30th May 2019 3
• In April of 2017, a global web of eight radio telescopes located in six places (Chile, Mexico, Spain, Hawaii, Arizona and the Antarctic), the collective network that makes up the EHT, began surveying the Messier 87 black hole, as well as the black hole at the center of our own Milky Way galaxy. → talk by S. Ohdachi
• An image is a 2D representation of electromagnetic waves either emitted or reflected by an object or set of objects
Infrared VUV Soft X-ray→ In this session
BöckenhoffPisano
Farley Ohdachi
Image processing starts with image acquisition and preparation
04.06.2019 4
Getting images
Acquiring images
Calibration of the images
Spatiotemporal interpretation
IAEA Meeting on Fusion Data Processing, Validation and Analysis, 30th May 2019
Image processing starts with image acquisition and preparation
04.06.2019 5
Getting images
Acquiring images
R.J. Maqueda and G.A. Wurden, Nucl. Fusion, 39 (1999) 629
M. W. Jakubowski, et al., Review of Scientific Instruments 89, 10E116 (2018)
Instrumentation of next generation devices becomes challenging due to steady-state requirements and harsh conditions (ITER)Combination of large depth of field and high resolution challenging
Image processing starts with image acquisition and preparation
04.06.2019 6
Getting images
Acquiring images
Calibration of the images
Spatiotemporal interpretation
W7-X
→ See talk by F. Pisano
Thermography Image Analysis Thermal Events Trackingand Characterization
Control, Safetyand Protection
IAEA Meeting on Fusion Data Processing, Validation and Analysis, 30th May 2019
In the case of infrared images by careful selection of observed wavelength range we can assure to observe an emission from surface of PFCs, which makes interpretation much easier
Image Processing in Fusion
04.06.2019 IAEA Meeting on Fusion Data Processing, Validation and Analysis, 30th May 2019 7
Processing of physical information
Analysis
Interpretation
Extraction of the information
Image Processing in Fusion
04.06.2019 IAEA Meeting on Fusion Data Processing, Validation and Analysis, 30th May 2019 8
Processing of physical information
Extraction of the information
• Typical issue with an image is that it represents 3D objects on a 2D plane.
• In real life our brain has 3D models, which help us to interpret partially stereoscopic images.
• In fusion devices we interpret a 2D image, where the models are either non-existing or not accurate enough.
• Several methods are developed to the extract structural information from a visual scene to understand phenomena and even control the discharge
Image Processing in Fusion
04.06.2019 IAEA Meeting on Fusion Data Processing, Validation and Analysis, 30th May 2019 9
Processing of physical information
Extraction of the information
COMPASS
P. Hacek, et al., W
DS’1
4 (2
01
4) 2
21
-22
6
W7-X
T. Szepsi, et al., EP
S 20
17
P5
.11
9
• Assumption of thin emissive surface layer. • Identification of plasma boundary based
on the visible radiation belt at the edge with help of either equilibrium reconstruction (COMPASS) or field line tracing (W7-X).
Image processing starts with image acquisition and preparation
04.06.2019 10IAEA Meeting on Fusion Data Processing, Validation and Analysis, 30th May 2019
S. Lisgo, et al., Jo
urn
al of N
uclear M
aterials 39
0–3
91
(20
09
) 10
78
–10
80
MAST
• For a quantitative comparison with the model, pixel-based tomographic inversion is used to reconstruct the Da emission profile in the poloidal plane.
• The algorithm is represented by Ax = b, where x is the poloidal emission profile and b is the camera data.
• A is the ‘geometry matrix’, which maps the line-of-sight camera view onto a poloidal mesh and is calculated using a ray tracing code.
Processing of physical information
Extraction of the information
Sparse modelling for tomographic reconstruction
• Expansion of the emission profile using orthogonal patters with the sparse modelling (e.g., L1 regularization) is promising for tomographic reconstruction even when the condition for the reconstruction is quite severe.
• Synthetic data analysis shows that, if we use two tangentially viewing camera data, island-like structure can effectively detected.
→ See talk S. Ohdachi
Reconstruction with L1
Assumed profile
W7-X
LHD
Processing of physical information
Extraction of the information
MARFE detection based on optical flow detection.
04.06.2019 IAEA Meeting on Fusion Data Processing, Validation and Analysis, 30th May 2019 12
Processing of physical information
Extraction of the information
• The optical flow is the pattern of apparent motion of objects caused by the relative motion between an observer and a scene.
𝑓𝑠 ⋅ 𝑣 + 𝑓𝑡 = 0
• Method allows for motion detection. The optical flow is defined as the ‘flow’ of grey values at the image plane and it is an approximation of the motion field, i.e. of the real motion of the object in the 3D scene, projected onto the image plane.
JET
T. Craciunescu, et al. Plasma Phys. Control. Fusion 56 (2014) 114006A. Murari, et al., IEEE TRANSACTIONS ON PLASMA SCIENCE, VOL. 38, NO. 12, DECEMBER 2010
Analysis
Analysis of Scrape-Off Layer Filament Properties Using Visual Camera Data
13
Camera frame Elzar tomographic inversion
Processing of physical information
Analysis
IAEA Meeting on Fusion Data Processing, Validation and Analysis, 30th May 2019
MAST
• Detection of filaments with neural networks from visual cameras has been developed at CCFE
• Application to experiments allows determining the statistics of the filaments with unique flexibility
• → see talk of Tom Farley
Extraction of the information
Interpretation of images
04.06.2019 IAEA Meeting on Fusion Data Processing, Validation and Analysis, 30th May 2019 14
Processing of physical information
Analysis
Interpretation
Tore Supra
• Interpretation of surface temperature and heat flux to the carbon plasma components deduced from infrared images requires detailed knowledge of surface emissivity and presence of so-called surface layers (redeposited material)
• Temporal evolution of the infrared images allowed to identify areas affected by surface layers A. Ali, et al., Nuclear Materials and Energy 19 (2019) 335–339
Extraction of the information
Information required to understand an infrared image
15
• Layers of information for interpratationof an image
• Emissivity ε of the target material (not constant in time)
• Surface layers (not constant in time)
• Reflections map (dependent on scenario)
• Target distance D
• Angle α with normal to target surface
• 3D world coordinates: X,Y,Z, ϕ, θ
LoS
ϕ, θ
X,Y,Z
α
D
ε
PFC ID
CAD view
Processing of physical information
Analysis
Interpretation
Extraction of the information
Image changes properties during thedischarge
Thermographic image from W7-X divertor
Surface layer detection
A. Ali, et al., Nuclear Materials and Energy 19 (2019) 335–339
W7-X
IAEA Meeting on Fusion Data Processing, Validation and Analysis, 30th May 2019
Surface layers obtained withTHEODOR – 2D heat flux code
Automatic processing of large amount of data
04.06.2019 17
Processing of physical information
Analysis
Interpretation
Extraction of the information
• Any diagnostic data can be converted into electrical signal stored for offline use
• Present and future magnetic fusion experiments need to deal with petabytes of data, which cannot be processed in a typical approach.
• An automatic data processing is required.
• A promising approach to deal with large amount of data is shown in J. Vega et al., Rev. Sci. Instruments 81, 023505
• Universal Multi Event Locator allows for automatic detection of events in time traces (e.g. saw teeth) or IR data (hot spots).
• Works also on images with reduced resolution.
Plasma parameter reconstruction from divertor heat load with sparse data
04.06.2019 18
• Reconstruction of magnetic equilibrium parameters from infrared data with neural networks.
• Due to small data amount training set supported by simulated images.
• → see talk D. Böckhenhoff• ITER requires methods to train neural
networks w/o experimental data (e.g. avoidance of disruptions)
IAEA Meeting on Fusion Data Processing, Validation and Analysis, 30th May 2019
Processing of physical information
Analysis
Interpretation
Extraction of the information
Image analysis requires sophisticated software
04.06.2019 19
• WEST, W7-X: THERMAVIP – A. Puig-Sitjes, et al., Fusion Science and Technology, 74 (2017) 1–2, 116–124• JET: JUVIL - V. Huber, et al., Fusion Engineering and Design123(2017)979–985
IAEA Meeting on Fusion Data Processing, Validation and Analysis, 30th May 2019
Operating System
Qt
Acquisition Processing
ThermaVIP SDK
StorageTriggers
CameraLocal
storageGPU NetworkIO board
Camera driver
CUDAIO board driver
GUI
Acquisition and analysis
OS
Qt
ThermaVIPSDK
Network
GUI
DisplayWEST, W7-X
JET
04.06.2019 20
Thank you