Recent work in music structure analysis has shown the potential of deep features to highlight the underlying structure of music audio signals. Despite promising results achieved by such representations, dealing with the inherent hierarchical aspect of music structure remains a challenging problem. Because different levels of segmentation can be considered as equally valid, specifically designed representations should be optimized to improve hierarchical structure analysis. In this work, unsupervised learning of such representations using a contrastive approach operating at different timescales is explored. The proposed system is evaluated on flat and multi-level music segmentation. By leveraging both time and the hierarchical organization of music structure, we show that the obtained deep embeddings can encode meaningful patterns and improve segmentation at various levels of granularity.