I have found that the kernel tricks give a T2/SPE contribution matrix that is compose of equal to the number of trained sample, unlike PCA. I need contribution of individual variable to a fault. If anyone who handled this issue could help me, would be really great.