The reliable detection of incipient bearing damage in electromechanical drive systems remains a persistent challenge in industrial practice, particularly when naturally developed faults must be distinguished from the healthy baseline using signals already present in standard drive cabinets. This thesis addresses that challenge through a structured, two-phase experimental study conducted on the Paderborn University KAt-DataCenter benchmark dataset, which provides synchronously recorded vibration and motor phase current signals across a controlled spectrum of artificial and naturally developed bearing damage states. In the first phase, a rich feature set is constructed from both signal modalities spanning time-domain statistics, Welch power spectral density descriptors, envelope spectrum amplitudes at bearing characteristic frequencies, and motor current sideband features and four classical machine learning classifiers (SVM, Random Forest, XGBoost, and k-NN) are systematically benchmarked under two evaluation protocols that differ in whether training and test data share the same damage origin.