Understanding the neurophysiological basis of meditation is crucial for enhancing its therapeutic applications. Electroencephalography (EEG) offers an objective and real-time means to assess meditative states, yet several challenges hinder its widespread applications: temporal confounding, limited labeled data, and the presence of noisy labels. This paper addresses these challenges through a multi-session EEG acquisition paradigm with intervals, and a novel three-stage self-supervised learning framework specifically designed for meditative EEG analysis. We built two datasets using the designed experimental paradigm. The self-supervised learning framework comprises: (1) general representation pre-training using large-scale EEG data from diverse tasks; (2) labeled contrastive learning on task-specific data to capture discriminative features; (3) subject-dependent fine-tuning with regularization to enhance noise robustness and personalization. Together, these innovations enable more accurate and generalizable classification of meditative states while addressing critical data quality issues. The proposed approach paves the way for scalable, objective meditation assessment and lays a foundation for future EEG-based mental wellness applications.