Identification and Classification of Bearings Based on Vibration Data

Sammanfattning

This thesis investigates the feasibility of identifying specific bearing models using vibration data.While vibration analysis is the “gold standard” for fault detection [1], its application for model-specific “kinematic fingerprinting” is limited by a fundamental Physics-Data Gap. Using a two-stageapproach comparing deterministic kinematic matching against deep learning (CNN) architectures [2]this research attempts to map vibration features to SKF catalogue designations. Results demonstrate that identification is significantly hindered by a substantial, yet physicallyexpected, catalog collision rate. Even after grouping 34,629 entries into 12,963 unique kinematicfamilies, the global collision rate remains high at 94.02% [3]. Crucially, even when applying application-specific constraints to isolate only plausible candidates, a 78.79% collision rate persists. This suggeststhe “Collision Problem” is a fundamental reality of bearing kinematics and shared dimensional ratios,rather than an artifact of a broad, unconstrained search space. Furthermore, signal-to-noise ratio(SNR) thresholds in healthy components present a critical barrier. A central finding is the “Health-State Paradox,” where damage signatures are required to “illuminate” a bearing’s kinematic identity,but eventually mask it through signature drift [4]. The study concludes that vibration data alone isinsufficient for high-confidence identification, and recommends multi-modal sensor fusion as a viableindustrial path forward.