This study introduces an unsupervised classifier designed to analyze intracranial electroencephalography (iEEG) signals during sleep, aiming to uncover intrinsic patterns and transitions between sleep phases without predefined labels. Employing an iterative self-classification approach, the model begins with random label assignments to signal segments and refines both the model and labels through mutual feedback, enhancing the identification of homogeneous classes based on power spectra features. Dimensionality reduction via UMAP facilitates visualization of class distributions in latent space, revealing relationships among sleep states such as REM and non-REM. Results demonstrate channel-specific temporal classifications of sleep states, clustered spectral averages, and comparisons with established hypnograms, highlighting the tool's efficacy in characterizing brain activity variations and potential applications for detecting pathological alterations in sleep disorders.