High-dimensionality and nonlinear dynamic processes complexity of datasets acquired by distributed sensors, for the purpose of preventive maintenance assessment, in realistic complex systems, for example propulsion machinery, create significant challenges in feature extraction, decreasing the clarity to form a reliable and complete picture of damage indication if not processed holistically but at the single sensor level as classically and routinely done. To overcome limitations, we extract features by applying the unparalleled Advanced Proper Orthogonal Decomposition (APOD) to reduce big triaxial acceleration datasets associated with a drivetrain system used as a test bed. APOD demonstrates high and generalized performance, due to its capacity to identify the data geometry skeleton structure and thus revealing data patterns as modal structures developed in the high-dimensional space where the geometry of the dataset unfolds in all possible directions. Here, raw datasets are acquired by distributed triaxial sensors during the operation of a typical drivetrain system. These tensor-valued datasets are decomposed using the APOD algorithm and the norms of the POD modes shapes are computed. The set of the normalized POD modal vectors are mapped into a characteristic cluster of data points sitting in the 3-dimensional Euclidean space. The graph connecting this cluster of points is the claimed extraction of features from a big dataset since the former is the most optimum reduction of the big dataset. This feature is referred to as the characteristic scaffold of the tensor dataset. The major contribution of this work is the derivation of a damage diagnosis spectrum index (DDSI) distributed over the wave number of the computed POD modes. This is derived by a geometry deformation operation applied on the characteristic scaffold as it suffers a damage-inflected variation, or deformation, from a reference healthy state to a non-healthy state, that is the damaged one.