I can't see anything obviously wrong, but here are a few things you could check:

(1) The option '-exp' in 3dcalc should be '-expr';
(2) Do you only have ONE run of original dataset with 554 TR's? I mean, did you run 3dDeconvolve with option '-concat'?
(3) Make sure those three subbricks [0], [2], and [50] are really the corresponding coefficients.
(4) You may consider adding a step function in the expression so that some noisy voxels outside the brain could be filtered out, which could be exactly the reason why you got those huge numbers of percent signal change (by the way, you could use ANFI to locate some of the voxels with absurdly huge % signal change). Do something like this:

3dcalc \
-a Decon${subject}_18view554Tr30SAN+orig[0] \
-b Decon${subject}_18view554Tr30SAN+orig[2] \
-c Decon${subject}_18view554Tr30SAN+orig[50] \
-d Decon${subject}_18view554Tr30SAN+orig[Number] \ ## put in the number for the whole-F subbrick
-expr "100*c/(a+b*(554/2))*step(d-F_value)" \ ##Use AFNI to find out a threshold F-value so that those significantly inactive voxels would be ignored during the percent signal conversion process
-fscale \
-prefix ${subject}_18view_Prcnt_AvN

(5) Personally I prefer normalizing the original dataset before running 3dDeconvolve. Check out Doug Ward's following message:

[afni.nimh.nih.gov]

Gang