Hi @shilin
Based purely on your code, I can’t really provide much direction. It seems that the methodology you have followed is valid. I cannot see where you are specifying the keepX range - this may have an impact. Look here for more info. If you are partioning your data into training and testing prior to the use of tune_diablo_ncomp(), then you are only tuning it on training data which is incorrect. The perf() function partitions the data for you within each repeat.
Apart from these suggestions, there isnt much else I can say. It may just be that your data cannot be discriminated on, or that your sample is not representative of the population you are examining.