Teaching Large Language Models to Express Knowledge Boundary from Their Own Signals