Page 196 - Special Topic Session (STS) - Volume 3
P. 196

STS540 Zhi Song et al.
                      Generally,  the  ARL  is  the  widely  used  criterion  for  assessing  the
                  performance of monitoring schemes. However, since the FIR feature is more
                  meaningful for early detection, we use two other design criteria rather than
                  depending solely on the ARL performance measure. One design criterion is
                  cumulative unconditional false alarm probability (CUFAP), which is denoted by
                  CUFAP()=P(  ≤   |). Another  is  cumulative  unconditional  true  signal
                  probability (CUTSP),  denoted as  CUTSP()=P(  ≤   |). Thereafter, we
                  may  use  a  simple  algorithm  to  optimize  the  design  of  FIR-based  EWMA
                  schemes via the following steps:

                  Step-1: Given , , , and a target IC ARL ( ), say , compute  for a given
                                                               0
                  EWMA-LS scheme without the FIR feature. The abbreviation LS stands for the
                  Location-Scale statistic, namely, Lepage or Cucconi, with  = (LS|IC).
                                                                          0

                  Step-2: Let   ∈ [0,1] is a small positive proper fraction chosen a-priori by the
                  practitioner,  denotes  the  maximum  allowable  CUFAP  till -th  test  sample.
                  Compute CUFAP() for a pre-specified  ∈ N. If CUFAP() >   , increase 
                  by  a  margin  of  0.01  and  re-compute   and  CUFAP( ).  Note  that,  in
                                                             0
                  general, the CUFAP() is a decreasing function and the    is an increasing
                                                                            0
                  function of  given other charting parameters, namely, , , , remain same.

                  Step-3: Repeat the Step-2, unless, CUFAP() ≤   .

                  Step-4: Set the CUFAP() =   , and use the L obtained at the end of the
                  Step-3, say  , to determine dynamic FIR quotient. Using   and   determine
                                                                          0
                                                                                 0
                              0
                  the distance, say dLS  between   and UCL at the first test sample. Set FIR
                                                  0
                  feature ℎ  ∈ [0,1], where ℎ   = ℎ implies that the starting value will be reset
                           
                  at  0ℎ  =  + ℎ ×  .
                                     
                            0

                  Step-5:  We  see,  in  general,  for  fixed   ,  if    increases  to   ℎ > 0, ,
                                                             0
                                                                                    0ℎ
                                                                    0
                  CUFAP() increases. Therefore, to maintain the specified preapproved level
                  of CUFAP(), we need to increase the  to   and consequently, the  ,
                                                                                           0
                                                               ℎ
                  will also increase. Compute   for each  ℎ =  0.01(0.01)1.00.
                                              ℎ
                                                         0ℎ

                  Step-6:  Given  , , ,   and  target  shift   =  ,  =     compute
                                                                          
                  CUTS      () for all ℎ  =  0.00(0.01)1.00.
                         ℎ,  , 

                  Step-7: Select dynamic FIR as ℎ   = ℎ for which CUTSP ℎ,  ,  () is maximum.
                     Because of the distribution-free nature of the proposed class of monitoring
                  schemes when the process is IC, determination of   value based on a fixed
                                                                     ℎ
                  CUFAP() will be the same for all continuous distributions. We may generate
                                                                     185 |I S I   W S C   2 0 1 9
   191   192   193   194   195   196   197   198   199   200   201