{"title": "Neural Network Star Pattern Recognition for Spacecraft Attitude Determination and Control", "book": "Advances in Neural Information Processing Systems", "page_first": 314, "page_last": 322, "abstract": null, "full_text": "314 \n\nNEURAL NETWORK STAR PATTERN \n\nRECOGNITION  FOR  SPACECRAFT ATTITUDE \n\nDETERMINATION AND CONTROL \n\nPhillip Alvelda,  A.  Miguel San Martin \n\nThe Jet Propulsion  Laboratory, \n\nCalifornia  Institute of Technology, \n\nPasadena, Ca.  91109 \n\nABSTRACT \n\ncomputational  bottlenecks \n\nCurrently,  the  most  complex  spacecraft  attitude  determination \nand  control  tasks  are  ultimately  governed  by  ground-based \nsystems  and  personnel.  Conventional  on-board  systems  face \nsevere \nserial \nmicroprocessors operating on inherently parallel problems.  New \ncomputer architectures based on the anatomy of the human brain \nseem  to  promise  high  speed  and  fault-tolerant  solutions  to  the \nlimitations  of serial  processing.  This  paper  discusses  the  latest \napplications of artificial neural networks to the  problem of star \npattern recognition  for spacecraft attitude determination. \n\nintroduced \n\nby \n\nINTRODUCTION \n\nBy design, a conventional on-board microprocessor can perform only \none  comparison or  calculation at a  time.  Image or  pattern  recognition \nproblems involving large  template sets and  high  resolution  can require \nan  astronomical  number  of  comparisons  to  a  given  database.  Typical \nmission  planning and optimization tasks  require calculations involving \na  multitude of  parameters,  where  each element  has  an inherent degree \nof  importance,  reliability  and  noise. \nthe  most  advanced \nsupercomputers  running  the  latest  software  can  require  seconds  and \neven minutes to execute a  complex  pattern recognition or expert system \ntask, often providing incorrect or inefficient solutions to problems that \nprove  trivial  to ground control specialists. \n\nEven \n\nThe intent of ongoing research is to develop a  neural network based \nsatellite attitude determination system prototype capable of determining \nits  current  three-axis  inertial  orientation.  Such  a  system  that  can \ndetermine in real-time, which direction the satellite is  facing, is needed \nin  order  to  aim  antennas,  science  instruments,  and  navigational \nequipment.  For a  satellite to be  autonomous (an important criterion in \ninterplanetary  missions,  and  most  particularly  so  in  the  event  of  a \nsystem  failure),  this  task must be  performed in a  reasonable amount of \ntime  with  all  due  consideration  to  actual  environmental,  noise  and \nprecision constraints. \n\nCELESTIAL ATTITUDE  DETERMINATION \n\nUnder  normal  operating  conditions  there  is  a  whole  repertoire  of \nspacecraft systems  that operate in  conjunction to  perform  the  attitude \ndetermination  task,  the  backbone  of  which  is  the  Gyro.  But  a  Gyro \nmeasures only chaDles in orientation.  The current attitude is stored in \n\n\fNeural Network Star Pattern Recognition \n\n318 \n\nvolatile on-board  memory and  is  updated as  the ,yro system  inte,rates \nvelocity  to  provide chanle in  anlular position.  When  there  is  a  power \nsystem failure for any reason such as a sinlle-event-upset due to cosmic \nradiation, an currently stored attitude lafor.atloa Is  LOST! \n\nOne  very attractive way of recoverinl attitude information with no \na  priori knowledge  is  by USinl on-board  imalinl and computer systems \nto: \n\n1.) \n\nImage  a  portion of the  sky, \n\n2.) \n\n3.) \n\n4.) \n\n5.) \n\nCompare the characteristic pattern of stars in the sensor field(cid:173)\nof-view  to an on-board star catalog, \n\nThereby identify the stars in the sensor FOV [Field Of View], \n\nRetrieve  the  identified star coordinates, \n\nTransform  and  correlate  FOV  and  real-sky  coordinates  to \ndetermine spacecraft attitude. \n\nBut the problem of matching a  limited field  of view that contains a \n\nsmall  number  of stars  (out  of  billions and  billions of them),  to an  on(cid:173)\nboard  fUll-sky  catalol containing  perhaps  thousands  of stars  has  lonl \nbeen  a severe computational  bottleneck. \n\nD14~---------;\"':~~::.-r----\u00ad\n\nD13'---~='7\"\"T \n\n,; \n\n, ; '  \n\n, /   , \n/ \nPAIR  21  PAIR  703 \nPAIR  22  PAIR  704 \n\n\\ \n\n',STORED PAIR  ADDRESS \n\n,'\" \nPAIR  70121 \nPAIR  70122 \n\nGEOMETRIC \nCONSTRAINTS \n\nFicuN I.) Serial .tar I.D.  catalol rorma'  and  rnethodololY. \n\nThe \n\nlatest  serial  allorithm \n\nrequires \napproximately  650  KBytes  of  RAM  to store  the  on-board  star catalol. \nIt  incorporates  a  hilhly  optimized  allorithm  which  uses  a  motorola \n68000 to search a sorted database of more than 70,000 star-pair distance \nvalues  for  correlations  with  the  decomposed  star  pattern  in  the  sensor \nFOV.  It performs the  identification  process  on  the  order of  I  second \n\nto  perform \n\ntask \n\nthis \n\n\f316 \n\nAlvelda and  San Martin \n\nwith a  success  rate of 99  percent.  But  it does  Dot  fit  iD  the spacecraft \noD-board  memory,  and  therefore,  no  such  system  has  flown  on  a \nplanetary spacecraft. \n\n\u2022  USES SUN  SENSOR AND ATTITUDE  MANEUVERS \n\nTO  SUN \n\nTO  SUN \n\nCANOPUS \n\nFicuN J.) Current Spacecraft attitude inrormation  recovery  lequence. \n\nAs  a  result,  state-of-the-art  interplanetary  spacecraft  use  several \nindependent sensor systems in ~onjunction to determine attitude with no \na  priori  knowledge.  First, the  craft is  commanded  to slew  until  a  Sun \nSensor  (aligned  with  the  spacecraft's  major  axis)  has  locked-on  to  the \nsun.  The  craft  must  then  rotate  around  that axis  until  an appropriate \nstar  pattern at  approximately  ninety  degrees  to  the  sun  is  acquired  to \nThe  entire  attitude \nprovide \nacquisition sequence  requires  an  absolute  minimum  of  thirty  minutes, \nand  presupposes  that all  spacecraft actuator and  maneuvering systems \nare  operational.  At  the  phenomenal  rendezvous  speeds  involved  in \ninterplanetary  navigation,  a  system  failure  near  mission  culmination \ncould mean an almost complete loss of the most  valuable scientific data \nwhile  the spacecraft performs its initial attitude acquisition sequence. \n\nthree-axis  orientation  information. \n\nNEURAL  MOTIVATION \n\nThe parallel architecture and collective computation properties of a \nneural  network  based  system  address  several  problems  associated  with \nthe  implementation  and  performance  of  the  serial  star  ID  algorithm. \nInstead  of  searching  a  lengthy  database  one  element  at  a  time,  each \nstored  star  pattern  is  correlated  with  the  field  of  view  concurrently. \nAnd whereas standard memory storage technology  requires one address \nin RAM per star-pair distance, the neural star pattern representations are \nstored  in  characteristic  matrices  of  interconnections  between  neurons. \nThis distributed data set representation has several desirable properties. \nFirst of all, the 2N  redundancy of the serial star-p.air scheme (i.e. which \nstar is at which end of a  pair) is  discarded and a  new more  compressed \nrepresentation emerges from  the neuromorphic architecture.  Secondly, \nnoise,  both  statistical  (i.e  thermal  noise)  and  systematic  (i.e.  sensor \nlimitations),  and  pattern  invariance  characteristics  are \nprecision \n\n\fNeural Network Star Pattern Recognition \n\n31 7 \n\nincorporated  directly  into  the  preprocessing  and  neural  architecture \nwithout extra circuitry. \n\nThe first  neural approach \n\nThe  primary  motivation  from  the  NASA  perspective  is  to  improve \nsatellite attitude determination performance and enable on-board system \nimplementations.  The problem methodology for the neural architecture \nis  then slightly different than  that of the serial  model. \n\nInstead of  identifying every  detected  st~r in  the  field  of  view,  the \nneural system identifies a single 'Guide Star' with respect to the pattern \nof dimmer stars around it, and correlates that star's known position with \nthe sensor FOV to determine the pointing axis.  If needed, only one other \nstar  is  then  required  to  fix  the  roll  angle  about  that axis.  So  the  core \nof  the  celestial attitude  determination  problem changes  from  multiple \nstar identification and correlation, single star pattern  identification. \n\nThe  entire  system  consists  of  several  modules  in  a  marriage  of \ndifferent technologies.  The first neural system architecture uses already \nmature(i.e.  sensor/preprocessor) technologies  where  they  perform  well, \nand  neural \ntechnology  only  where  conventional  systems  prove \nintractable.  With an eye towards rapid prototyping and implementation, \nthe  system  was  designed  with  technologies  (such  as  neural  VLSI)  that \nwill be available in less  than one  year. \n\nSYSTEM ARCHITECTURE \n\nThe Star Tracker sensor system \n\nThe  system  input  is  based  on  the  ASTROS  II  star  tracker  under \ndevelopment in  the Guidance and Control section at the Jet Propulsion \nLaboratory.  The Star tracker optical system images a defocussed portion \nof the sky (a star sub-field) onto a charged coupled device (C.C.D.).  The \ntracker  electronics  then  generate  star  centroid  position  and  intensity \ninformation and  passes  this list  to the  preprocessing system. \n\nThe Preprocessln8 system \n\nThis centroia ind intensity information is passed to the preprocessing \nsubsystem where  the star pattern is  treated to extract noise and pattern \ninvariance.  A  'pattern field-of-view' is defined as centered aroun~ ~he \nbrightest (Le.  'Guide Star') in the central portion of the sensor field-of(cid:173)\nview.  Since  the pattern  FOV  radius is  one half that of the sensor  FOV \nthe pattern for that 'Guide Star' is  then based on a  portion of the image \nthat  is  complete,  or  invariant,  under  translational  perturbation.  The \npreprocessor  then  introduces  rotational  invariance  to  the  'guide-star' \npattern by  using only  the distances of all other dimmer stars inside  the \npattern FOV  to  the central  guide star. \n\nThese  distances  are  then  mapped  by  the  preprocessor  onto  a  two \ndimensional  coordinate  system  of  distance  versus  relative  magnitude \n(normalized  to  the guide star, the brightest star in  the Pattern FOV)  to \nbe  sampled  by  the  neural  associative  star  catalog.  The  motivation  for \nthis  distance  map  format  become  clear  when  issues  involving  noise \ninvariance and  memory  capacity are  considered. \n\n\f318 \n\nAlvelda and San Martin \n\nBecause the ASTROS Star Tracker is a  limited precision instrument, \nmost  particularly  in  the  absolute  and  relative  intensity  measures,  two \nmajor  problems  arise.  First,  dimmer  stars  with  intensities  near  the \nbottom of the dynamic range of the C.C.D.  mayor may  not  be  included \nin the star pattern.  So, the entire distance map is scaled to the brightest \nstar  such  that  the  bright,  high-confidence  measurements  are  weighted \nmore  heavily, while  the dimmer and  possibly  transient stars are  of less \nimportance  to  a  given  pattern.  Secondly,  since  there  are  a  very  large \nnumber  of  stars  in  the  sky,  the  uniqueness  of  a  given  star  pattern  is \ngoverned  mostly  by  the  relative star  distance  measures  (which,  by  the \nway,  are  the  highest  precision  measurements  provided  by  the  star \ntracker). \n\nIn addition, because of the limitations in expected neural hardware, \na  discrete  number  of  neurons  must  sample  a  continuous  function.  To \nretain  the  maximum  sample  precision  with  a  minimum  number  of \nneurons, the neural system uses the biological mechanism of a  receptive \nfield  for  hyperacuity.  In other words, a  number of neurons  respond  to \na  single distance stimulus.  The process  is analogous to that used on the \ndefocussed  image of a  point source on  the C.C.D.  which was  integrated \nover  several  pixels  to  generate  a  centroid  at  sub-pixel  accuracies.  To \nrelax the demands on hardware development for the neural module, this \npoint  smoothing  was  performed  in  the  preprocessor  instead  of  being \nintroduced  into  the  neural  network  architecture  and  dynamics.  The \nequivalent neural  response  function  then becomes: \n\nX\u00b7 I \n\nN L 'l'k  e  -(Ili  -\n\nk=l \n\nIlle  )2/tl. \n\nwhere: \n\nXi \n\nN \n\nILi \n\nis  the sampling activity of neuron  i \n\nis the number of stars in the Pattern Field Of View \n\nis  the  position of neuron i  on  the sample axis \n\nILk  \u2022  is  the  position of the  stimulus from  star  k  on  the \n\nsample axis \n\nis the magnitude scale factor of star k, normalized \nto the brightest star in the PFOV, the 'Guide star' \n\nis  the width of the gaussian  point spread function \n\nThe Neural system \n\nThe neural system, a  106 neuron, three-layer, feed-forward network, \nsamples  the  scaled  and  smoothed  distance  map,  to  provide  an  output \nvector  with  the  highest  neural  output  activity  representing  the  best \nmatch  to  one  of  the  pre-trained  guide  star  patterns.  The  network \ntraining  algorithm  uses  the  standard  backwards  error  propagation \n\n\fNeural Network Star Pattern Recognition \n\n319 \n\nalgorithm  to  set  network  interconnect  weights  from  a  training  set  of \n'Guide  Star'  patterns  derived  from  the  software  simulated  sky  and \nsensor  models. \n\nSimulation  testbed \n\nThe  computer simulation  testbed  includes  a  realistic  celestial  field \nmodel,  as  well  as  a  detector  model  that  properly  represents achievable \nposition and intensity resolution, sensor scan rates, dynamic range, and \nsignal  to  noise  properties.  Rapid  identification  of  star  patterns  was \nobserved  in  limited  training sets  as  the  simulated  tracker was  oriented \nrandomly  within  the celestial sphere. \n\nPERFORMANCE RESULTS AND PROJECTIONS \n\nIn  terms  of  improved  performance  the  neural  system  was  quite  a \nsuccess,  but  not  however  in  the  areas  which  were  initialJy  expected. \nWhile a VLSI implementation might yield considerable system speed-up, \nthe  digital  simulation  testbed  neural  processing  time  was  of  the  same \norder as the serial algorithm, perhaps slightly better.  The success rate of \nthe  serial  system  was  already  better  than  99%.  The  neural  net  system \nachieved  an  accuracy  of  100%  when  the  systematic  noise  (i.e.  dropped \nstars) of the sensor was  neglected. \n\nWhen  the  dropped  star  effect  was  introduced,  the  performance \nfigure  dropped  to 94%.  It was  later discovered  that the  reason  for  this \n'low'  rate  was  due  mostly  to  the  limited  size  of  the  Yale  Bright  Star \ncatalog at higher  magnitudes (lower star brightness).  In  sparse  regions \nof the sky,  the  pattern  in  the sensor  FOV  presented  by  the  limited sky \nmodel  occasionally  consisted of only  two or  three  dim stars.  When  one \nor two of them drop out because of the Star sensor  magnitude precision \nlimitations.  at  times.  there  was  no  pattern  left  to  identify.  Further \nexperiments  and  parametric  studies  are  under  way  using  a  more \ncomplete Harvard Smithsonian catalog. \n\nThe big gain was in terms of required memory.  The serial algorithm \nstored over  70,000 star pairs at high  precision  in addition to code  for a \nrather complex  heuristic, artificial intelligence type of algorithm for  a \ntotal  size  of  650  KBytes.  The  Neural  algorithm  used  a  connectionist \ndata  representation  that  was  able  to  abstract  from  the  star  catalog, \npa ttern  class  similarities,  orthagonalities,  and  in variances  in  a  highly \ncompressed  fashion. \nNetwork  performance  remained  essentially \nconstant until interconnect precision was decreased to less than four bits \nper synapse.  3000 synapses at four  bits per synapse  requires very  little \ncomputer memory. \n\nThese simulation results were all derived  from a  monte carlo run of \n\napproximately  200,000 iterations using  the  simulator testbed. \n\n\f320 \n\nAlvelda and San Martin \n\nCONCLUSIONS \n\nBy  means  of  a  clever  combination  of  several  technologies  and  an \nappropriate  data  set  representation,  a  star  10 system  using  one  of  the \nmost  simple  neural  algorithms  outperforms  those  using  the  classical \nserial ones  in several aspects, even while  running a  software simulated \nneural network.  The neural simulator is approximately ten times faster \nthan  the equivalent serial algorithm and  requires less  than one seventh \nthe  computer  memory.  With  the  transfer  to  neural  VLSI  technology, \nmemory requirements will virtually disappear and processing speed will \nincrease  by at least an order of magnitude. \n\nW1)ere  power and weight requirements scale with the hardware chip \ncount, and every pound that must  be  launched into space costs millions \nof  dollars,  neural  technology  has  enabled  real-time  on-board  absolute \nattitude  determination  with  no  a  priori \nthat  may \neventually make several accessory satellite systems like horizon and sun \nsensors  obsolete,  while  increasing  the  overall  reliability  of spacecraft \nsystems. \n\ninformation, \n\nAckaowledgmeats \n\nWe  would  like  to  acknowledge  many  fruitfull conversations  with C.  E. \nBell, J.  Barhen and S.  Gulati. \n\nRefereaces \n\nR.  W.  H.  van  Bezooijen.  Automated  Star  Pattern  Recognition  for  Use \nWith  the  Space  Infrared  Telescope  Facility  (SIRTIF).  Paper  for \ninternal  use at The Jet  Propulsion  Laboratory. \n\nP.  Gorman,  T.  J.  Sejnowski.  Workshop  on  Neural  Network  Devices and \nApplications (Jet Propulsion Laboratory, Pasaden, Ca.) Document 0-\n4406,  pp.224-237. \n\nJ.  L.  Lunkins.  Star  pattern  Recognition  for  Real  Time  Attitude \n\nDetermination.  The Journal of Astronautical Science(l979). \n\nD.  E.  Rummelhaft,  G.  E.  Hinton.  Parallel  Distributed  Processing,  eds. \n(MIT Press, Cambridge, Ma.)  Vol.  1 pp.  318-364. \n\nP.  M  Salomon,  T.  A.  Glavich.  Image  Signal  Processing  and  Sub-Pixel \n\nAccuracy Star Trackers.  SPfE  vol.  252  Smart Sensors  II  (1980). \n\n\fNeural Network Star Pattern Recognition \n\n321 \n\nC.C.D .  Image \n\nPreprocessor \n\nDistance \nMap \n\nNeural \nSampler \n\nNeural \nOutput \n\nStar Attitude \nLook-up Table \n\nRedtus  from Guide Ster \n\naa  DO  a  cD DOc a  CDc \nI I  I  I \n\nI  I  I  I \n\nI \n86 \n\n18 \n\n28 \n\n41  50 \n\n71 \n\n32 \n\n30 \nSt  1/1  R.A.  Dec. \n27  -1.3  2.45 \n:::::~f;r }=:=n:, :::~J;=r:::= \n0.2  0.68 \n\n29 \n\n\f322 \n\nAlvelda and San Martin \n\nPROTOTYPE  HARDWARE  IMPLEMENTATJON \n\nSERIAL PR')CESS()R \nCCMROlLER PNO \nSIC INTERFACE \n\nTOACS \n\n: ....................................................................................................................................................................... : \n\nq  q  Cl.JQ \n\n80 \n\nq  q  ~Q \n\nNEURAL \n\nPROCESSOR \n\n\u2022\u2022\u2022\u2022\u2022\u2022 \n\n1 \n\nCORRELATOR \n\n............................................................................................................................................................................................................................... \n\n\f", "award": [], "sourceid": 177, "authors": [{"given_name": "Phillip", "family_name": "Alvelda", "institution": null}, {"given_name": "A.", "family_name": "San Martin", "institution": null}]}