Manufacturing industries now leverage high-dimensional streaming video data from diverse sensors, represented as tensors (multidimensional arrays of channels × signals × time), for real-time monitoring, inspection, and quality control; however, this data often contains redundancy and captures only a subset of the complete dataset. Selecting effective dimensionality reduction and feature extraction methods for high-dimensional data structures remains challenging. To address these challenges, this paper presents a comparative framework for effective dimensionality reduction and feature extraction, utilizing supervised methods—Principal Component Analysis (PCA) and Independent Component Analysis (ICA)—alongside the unsupervised Multilinear-PCA (MPCA), which can more effectively handle multidimensional tensor structures compared to the 1-D or 2-D limitations of PCA and ICA. We evaluate this comparative framework for classifying fabric design patterns using high-dimensional video data captured from various fabric surface weave patterns. The videos are converted into sequential RGB frames and analyzed using the Gray-Level Co-occurrence Matrix (GLCM) for feature extraction, after which the dimensionality of the GLCM features is reduced with PCA, ICA, and MPCA, and the features are classified using supervised machine learning techniques for fabric design pattern recognition. MPCA achieves a 0.022% dimensionality reduction by extracting uniformly distributed features that effectively capture correlated fabric design patterns, unlike the less organized distributions from PCA and ICA. The fabric pattern classification accuracy achieved with MPCA, PCA, and ICA was 99.02%, 95.21%, and 92.68%, respectively. These results suggest that the proposed framework effectively facilitates dimensionality reduction and feature extraction in both supervised and unsupervised methods for high-dimensional video data.
