A hybrid deep-clustering framework is introduced for unsupervised pattern recognition in complex, high-dimensional datasets. Representation learning is coupled with structure-aware clustering in a mutually reinforcing loop. A contrastive encoder equipped with a sparsity-regularised bottleneck produces noise-robust embeddings, while an adaptive clustering head alternates between density- and graph-based objectives to capture both manifold and community structure. Pseudo-labels from the clustering head refine the encoder through consistency and neighbourhood-preservation losses, and a temperature-controlled target distribution stabilises assignment updates. Real-world imperfections are addressed via missing-value masking, augmentation-invariant training, and a curriculum that gradually increases cluster granularity. Computational complexity is controlled by mini-batch graph construction and linear-time neighbourhood approximations, enabling scalability to millions of samples. Across heterogeneous benchmarks, image datasets, tabular sensor logs, and text embeddings, the approach improves normalised mutual information and clustering accuracy while reducing collapse and initialisation sensitivity. Ablation studies verify the contribution of each component, and sensitivity analyses indicate robustness to noise ratio, class imbalance, and unknown cluster counts. These results suggest that tightly coupling representation learning with complementary clustering objectives yields reliable and scalable unsupervised pattern discovery in challenging data.