Advanced methods for 3D magnetic localization in industrial...
Transcript of Advanced methods for 3D magnetic localization in industrial...
-
Abhaya Chandra Kammara S. and Andreas König
Institute of Integrated Sensor Systems
Department of Electrical and Computer Engineering
Advanced methods for 3D magnetic localizationin industrial process distributed data-logging
with a sparse distance matrix
Abhaya Chandra Kammara S. & Prof.Dr.-Ing. Andreas König
-
Abhaya Chandra Kammara S. and Andreas König
Overview
1. Introduction2. State of the art3. Localization setup4. Localization algorithms5. Results6. Conclusion
-
Abhaya Chandra Kammara S. and Andreas König
Introduction
Goal: Localization in industrial containers filled by liquidor other media and in smart environments
Problem: Traditional techniques using RF, light based and ultrasound based do not work
Reason:➔ High attenuation of signals in liquids ➔ High reflectivity in containers
-
Abhaya Chandra Kammara S. and Andreas König
Localization
Solution: Magnetic Localization
Advantages: ➔Attenuation does not vary in liquids.➔Is not reflected by containers.
Applications:➔Head Tracking [3]➔Military[14]➔Silent localization of underwater sensors[8]➔Location and orientation tracking[9]➔Medical Systems[13][12]
-
Abhaya Chandra Kammara S. and Andreas König
AMR Sensor
Among the various types of magnetic sensors, AMR sensor from Sensitec Gmbh was chosen for this project. A 3D version of the sensor was designed in ISE lab.
-
Abhaya Chandra Kammara S. and Andreas König
Localization - Setup
Artificial magnetic fields are generated by circular coils placed around the container. Distance is calculated from each coil (2 phases), only one coil is on at one point of time.
The image shows the setup of industrial campaign, This data collected by Stefano Carrella and Dennis Groben is used as benchmark data in this work.
-
Abhaya Chandra Kammara S. and Andreas König
Localization - Setup
Distance calculation The distances are calculated from the voltages obtained from 3D AMRSensor using the following formulae from [18]
Where,
S is the sensitivity, VS is sensor voltage, G is the gain of the amplifier, n is the number of windings, R denotes the radius of the coil and μ0 = 4 × π × 1e−7 .
-
Abhaya Chandra Kammara S. and Andreas König
Localization Algorithms
MDS Mapping
Once the distances are obtained, We can localize the sensor using any range based localization techniques
●Multidimensional scaling techniques are popular methods for range based localization.
●We are interested in Non linear mapping(Sammon's mapping). ●The localization is done based on reducing the error function,
where N is number of nodes, dij is Euclidean distance between Xi and Xj inHigher dimension and dij is Euclidean distance between Yi and Yj in lowerdimension.
-
Abhaya Chandra Kammara S. and Andreas König
Localization Algorithms
NLMR:●No inter sensor distances are present in our problem.●NLM is simplified to NLMR with this modified error function mentioned in [5]
Where,
dXij is the distance between recall and training data in high dimensional Space and K is the number of code book vectors.
-
Abhaya Chandra Kammara S. and Andreas König
Localization Algorithms
NLMR Gradient descent
Where yiq (m + 1) is the new position, MF is the magic factor which reduces With time, yiq (m) is the current position and E(m) is the cost function at the current position.
With,
In the Gradient Descent approach we make use of the NLMR cost function and the following equations mentioned in [5].
-
Abhaya Chandra Kammara S. and Andreas König
Localization Algorithms
Modification to NLMR for localization
●The specialty of NLMR approach is that it matches perfectly with our requirements. ●We do not obtain an inter-point distance between the nodes. ●We have a set of mapped positions (coils).●We have to make some changes to the original algorithm to make it work for localization.●In our case we do not require a dimensionality reduction. So the known locations(magnetic coils) will act as the trained data set. ●We do not have any location for the unknown value, however we have the distance information from the known locations which can be directly given to the algorithm.●Unlike Sammon’s mapping we do not have to do a reverse mapping(conformal transform) after we get our output.
-
Abhaya Chandra Kammara S. and Andreas König
Localization Algorithms
NLMR Simulated annealing
We use the basic simulated annealing We start with a high temperature which is reduced over the number of cycles (here 20) and reduce the chances of getting a worse solution as the temperature decreases. An energy component is not explicitly used in this approach. The algorithm runs for 500 iterations to get the best solutions. The number of iterations required was found heuristically.
-
Abhaya Chandra Kammara S. and Andreas König
Localization Algorithms
NLMR PSO
Standard particle swarm optimization described in [4] is used with C1 = C2 = 1 and the algorithm has no inertia. Having no inertia helped in faster convergence of the algorithm. 40 particles were used with 150 generations found the best results in the experiments.The standard PSO was unable to converge to a solution using Sammon’s mapping and required special approaches to reach convergence. However, in the modified NLMR method the standard PSO is able to converge easily.
-
Abhaya Chandra Kammara S. and Andreas König
Localization Algorithms
Multilateration
Multilateration is used in wireless sensor networks and a standard technique for efficient and effective localization. It serves as a reference in our work. Traditionally the time difference of arrival of the signal is used in multilateration technique, We use the distances obtained with our magneticlocalization system. In our approach we make use of the linear least squaresmethod (Moore-Penrose pseudo-inverse) to get the best fit solution
-
Abhaya Chandra Kammara S. and Andreas König
Experimental setup- Industrial data
Location of ground truth and anchors
trials pos trials/pos coils Coil dia windings V-sensor I-coils325 40 2 to 20 120 0.25m 230 3.7V 3A
-
Abhaya Chandra Kammara S. and Andreas König
Experimental results
All values are in metres
-
Abhaya Chandra Kammara S. and Andreas König
Experimental results
All values are in metres
All Coils NLMR-GD NLMR-SA NLMR-PSO Multilateration
LEμ 0.17951 0.17914 0.11429 0.16702
LEσ 0.13097 0.12557 0.0658 0.15194
Max LE 0.76031 0.62029 0.69258 1.31961
Min LE 0.0223 0.0172 0.00911 0.01302
μX -0.0935 -0.0978 -0.00121 -0.05915
μY 0.08262 0.08639 0.01846 -0.02365
μZ 0.0261 0.02201 0.02336 -0.0019
σ X 0.14338 0.1349 0.06336 0.12351
σ Y 0.08881 0.08591 0.07171 0.11815
σ Z 0.06942 0.06904 0.08571 0.13306
Max err X 0.67247 0.5581 0.50856 0.48092
Max err Y 0.58066 0.55335 0.50618 0.60638
Max err Z 0.20882 0.21477 0.3002 1.15277
Min err X 0.00017 0.00001 0.00018 0.00043
Min err Y 0.0014 0.0006 0.00065 0.0002
Min err Z 0.00026 0.00026 0.00021 0.00011
Table data of chart in previous slide
-
Abhaya Chandra Kammara S. and Andreas König
Experimental results – Coil selection
Variations between expected distancesand acquired distances for 325 experiments
1) We can observe that the error in coil 12 is much lower than error in coil 3 even though coil 3 is much closer to coil 12. 2)The heuristic of increasing variations with distances does not hold water.3) Angle plays a major role in distance errors. (Angle was disregarded in the distance calculation formulae- An alternative solution was proposed in [18])
-
Abhaya Chandra Kammara S. and Andreas König
Experimental results -Selected coils
Coils selected – 1,2,4,5,6,7,8,9
-
Abhaya Chandra Kammara S. and Andreas König
Experimental results -Selected coils
Sel. Coils NLMR-GD NLMR-SA NLMR-PSO MultilaterationLE μ 0.10106 0.09345 0.09534 0.13912
LE σ 0.05822 0.05861 0.06129 0.11206
Max LE 0.6943 0.6837 0.70647 0.80797
Min LE 0.01774 0.01126 0.01239 0.00573
μX -0.00238 0.01458 0.00813 -0.03728
μY 0.00126 0.0015 0.00686 -0.0376
μZ 0.00323 0.02655 -0.00776 0.03375
σ X 0.0497 0.04145 0.04564 0.07312
σ Y 0.06513 0.0677 0.0629 0.09209
σ Z 0.0829 0.07033 0.08144 0.11892
Max err X 0.19471 0.1339 0.17319 0.19803
Max err Y 0.5235 0.53103 0.49736 0.49735
Max err Z 0.508 0.57788 0.58183 0.78424
Min err X 0.00043 0.0001 0.00043 0.00012
Min err Y 0.00003 0.00007 0.00065 0.00003
Min err Z 0.00033 0.00021 0.00009 0.00018
-
Abhaya Chandra Kammara S. and Andreas König
Experimental results – ISE lab
trials pos trials/pos coils Coil dia windings V-sensor I-coils56 56 1 6 0.13m 100 5V 5A
All positions are on one quadrant of the XY region. There was no variation in Z for ground truth positions
-
Abhaya Chandra Kammara S. and Andreas König
Experimental results – ISE lab
NLMR-GD NLMR-SA NLMR-PSO Multilateration
LE μ 0.07169 0.07137 0.08715 0.12526
LE σ 0.03312 0.03429 0.03428 0.07147μX -0.05512 0.05434 0.05112 0.10509
μY 0.03315 0.03307 0.03383 -0.05274
μZ 0.00532 -0.00639 -0.00744 -0.00212
σ X 0.02857 0.02953 0.03727 0.07489
σ Y 0.02062 0.02057 0.02612 0.03253
σ Z 0.02881 0.02978 0.05372 0.01735
First cut results from ISE lab setup All values are in metres
-
Abhaya Chandra Kammara S. and Andreas König
Conclusion
1) An adaptation of NLMR for localization was presented.2) This was applied with Gradient Descent, Simulated annealing and PSO on industrial data and ISE lab data.3) The results were compared with Multilateration and obtained better results.4) Results were much better than conventional NLM from first cut analysis on the industrial data.5) The results of Simulated annealing and PSO were much better than gradient descent.6) The standard PSO was able to give very good results with all coils. The results are comparable to other methods using selected coils.
-
Abhaya Chandra Kammara S. and Andreas König
References
References1. H. P. Kalmus: A new guiding and tracking system, IRE Trans. Aerosp. Navig.Electron., vol. 9, pp. 710, 1962.2. J.W. Sammon: ”A nonlinear mapping for data structure analysis”, IEEE Transac-tions on Computers, C-18: 401-409, 19693. F. E. Raab et al.: Magnetic position and orientation tracking system, IEEE Trans.Aerosp. Electron. Syst., vol. 5, pp. 709717, 1979. ISSN 182378434. J. Kennedy, R.C. Eberhart, ”Particle swarm optimization”, Proc. IEEE Interna-tional Conference on Neural Networks, IJCNN, 19955. König, A.: Interactive Visualisation and Analysis of Hierarchical Neural Projectionsfor Data Mining. In IEEE TNN, Special Issue on Neural Networks for Data Miningand Knowledge Discovery, pp. 615 - 624, Vol. 11, No.3, May, 2000.
-
Abhaya Chandra Kammara S. and Andreas König
References
6. A. K onig: ”Dimensionality reduction techniques for interactive visualization, ex-̈ploratory data analysis, and classification”, Pattern Recognition in Soft ComputingParadigm, World Scientific, N.R. Pal (eds.), 2: 1-37, 2001.7. Eugene Paperno, Ichiro Sasada, and Eduard Leonovich A New Method for MagneticPosition and Orientation Tracking- - IEEE TRANSACTIONS ON MAGNETICS,VOL. 37, NO. 4, JULY 20018. E. A. Prigge: A positioning system with no line-of-sight restrictions for clutteredenvironments, Dissertation, Stanford University, 20049. J. Callmer, M. Skoglund, and F. Gustafsson : Silent localization of underwatersensors using magnetometers. EURASIP J. on Advances in Signal Processing, 201010. Jörg Blankenbach and Abdelmoumen Norrdine: Position Estimation Using Artifi-cial Generated Magnetic Fields, 2010 International Conference on Indoor Position-ing and Indoor Navigation (IPIN), 15-17 September 2010, Zürich, Switzerland
-
Abhaya Chandra Kammara S. and Andreas König
References
11. Jörg Blankenbach, Abdelmoumen Norrdine and Hendrik Hellmers:Adaptive SignalProcessing for a Magnetic Indoor Positioning System Geodetic Institute, TechnischeUniversit ̈at Darmstadt - Short paper IPIN12. Giuseppe Placidi*, Danilo Franchi, Alfredo Maurizi and Antonello Sotgiu: Reviewon Patents about Magnetic Localisation Systems for in-vivo Catheterizations INFMc/o Dept. of Health Sci., Uni. of LAquila,Via Vetoio Coppito 2, 67100 LAquila, Italy13. Ascension Technology Corporation Products Application: [Online]. http ://www.ascension−tech.com/medical/pdf /T rakStarW RT SpecSheet.pdf , checkedNov. 5 201214. Polhemus: [Online]. Available: http://www.polhemus.com/?page =M ilitaryW hyM agneticT racking,checked Nov. 5 201215. K. Iswandy and A. König: Soft-Computing Techniques to Advance Non-LinearMappings for Multi-Variate Data Visualization and Wireless Sensor Localization.e-Newsletter IEEE SMC Soc., Issue #29, Dec. 2009.
-
Abhaya Chandra Kammara S. and Andreas König
References
16. Stefano Carrella, Kuncup Iswandy, Kai Lutz, and Andreas König: 3D-Localization of Low-Power Wireless Sensor Nodes Based on AMRSensors in Industrial and AmI Applications, VDE Verlag GmbH Berlin Offenbach, pp. 522-529, 2010.17. Kuncup Iswandy, Stefano Carrella, and Andreas König: Localization System for Low Power Sensor Nodes Deployed in Liquid-Filled Industrial Containers Based on Magnetic Sensing Tagungsband XXIV Messtechnisches Symposium des AHMT, 23.-25. September, pp. 108-121, 201018. Stefano Carrella, Kuncup Iswandy, and Andreas König: A System for Localization of Wireless Sensor Nodes in Industrial Applications Based on Sequentially Emitted Magnetic Fields Sensed by Tri-axial AMR Sensors 19. Stefano Carrella, Kuncup Iswandy, and Andreas König: System for 3D Localization and Synchronization of Embedded Wireless Sensor Nodes Based on AMR Sensors in Industrial Environments, Proceedings Sensor+test 201120. Sebastian Reinecke, Uwe Pöpping und Uwe Hampel: Autonome Sensorpartikel zur raumlichen Parametererfassung in groskaligen Behältern, Sensor & Test 201221. Pending Patent Application: Method and Apparatus for Determining the Spatial Coordinates of at least one Sensor Node in a Container, filing date: 18.05.2010, Int. Pub. 24.11.2011, Int.No.: WO 2011/144325 A2
Institute of Integrated Sensor Systems Department of Electrical and Computer EngineeringOverviewPowerPoint-PräsentationSlide 4Slide 5Slide 6Slide 7Slide 8Slide 9Slide 10Slide 11Slide 12Slide 13Slide 14Slide 15Slide 16Slide 17Slide 18Slide 19Slide 20Slide 21Slide 22Slide 23Slide 24Slide 25Slide 26Slide 27