To develop a new questionnaire for measuring lifeguards' vigilance, after gathering data from different literature, I found 20 scales contributing to lifeguards' vigilance. Then I started to design questions for each scale in the way that in each scale, there are some questions (4 to 10 questions) with different face but with the same concept. Finally I have come up to 142 items. Now I want to perform factor analysis for construct validity. I don't know if I could perform FA at the scale level instead of item level. (Scales and items in each scale have face validity too.)
You should perform factor analysis at the item level. In other words, you should input all of the items into a single factor analysis. This is assuming, of course, that you have enough cases (500+?). This would allow you to know what items measure what different constructs. Factor analysis at the scale level would be possible, but quite unorthodox. This might work if you have too few cases. This would assume that items within you scales all measure the same thing, which they likely do not, and this is what factor analysis is designed to test.
My understanding is that confirmatory factor analysis (CFA) is typically run separately for each scale in order to verify the expected factor structure of each. If your intent is to use these scales as IVs and DVs in an analysis (e.g., multiple regression), I'd use this route. However, if your intention is to perform an exploratory factor analysis on all of your items across scales and search for "emergent" factors not intended by the original authors, I think an argument could be made for the latter approach.