On the Intrinsic Self-Correction Capability of LLMs: Uncertainty and Latent Concept