@article{4755,
  author = {Nguyen Minh Tuan},
  title = {Optimizing High-Dimensional Data Analysis Through Latent Representation Learning, Contrastive Embedding, and Consensus Clustering for Sensor Intelligence},
  journal = {Digital Signal Processing and Artificial Intelligence for Automatic Learning},
  year = {2026},
  volume = {5},
  number = {2},
  doi = {https://doi.org/10.6025/dspaial/2026/5/2/51-79},
  url = {https://www.dline.info/dspai/fulltext/v5n2/dspaiv5n2_1.pdf},
  abstract = {High-dimensional sensor data present substantial analytical challenges due to feature sparsity, nonlinear
dependencies, noisy observations, and unstable clustering behavior in conventional machine learning
frameworks. Although deep representation learning and subspace clustering methods have significantly
advanced high-dimensional analytics, many existing approaches remain fragmented, operating independently
across latent representation learning, semantic embedding optimization, and clustering stages. This study
proposes a unified multi-stage unsupervised sensor intelligence framework integrating Variational
Autoencoder (VAE)-based probabilistic latent representation learning, contrastive representation
optimization using Triplet Networks, SimCLR, and BYOL, and consensus hybrid clustering for robust subgroup
discovery in high-dimensional digital sensor signals. The proposed framework is evaluated using the IEEE
DataPort binary-classification digital sensor signal dataset comprising 11 numerical sensor features
characterized by temporal dependencies and event-driven activation patterns. Experimental evaluation
incorporates quantitative clustering metrics including Silhouette Score, Davies–Bouldin Index, Calinski–
Harabasz Index, reconstruction analysis, cluster stability evaluation, comparative baselines, and consensus
clustering validation. Results demonstrate that probabilistic latent manifold learning substantially improves
structural organization, semantic continuity, and anomaly-sensitive clustering compared with conventional
dimensionality reduction and raw-feature clustering approaches. Hierarchical clustering provides balanced
subgroup segmentation, whereas spectral clustering exhibits strong sensitivity toward sparse anomalous
structures. The proposed consensus hybrid clustering mechanism combines these complementary strengths
to produce stable and semantically coherent latent partitions. The study contributes a robust and scalableframework for intelligent sensing, anomaly detection, cyber-physical monitoring, and unsupervised IoT
analytics.},
}