Bearings are critical to promote efficient and smooth mobility. Traditional fault diagnosis methods exhibit limitations in accuracy and adaptability facing noise and weak fault signatures. To address this, the study applies recursive quantification analysis (RQA) to diverse operational data and employs data reconstruction techniques. RQA extracts discriminative recurrence features characterizing system dynamics, while data reconstruction generates enhanced representations. Vibration signals are preprocessed using envelope analysis and filtering, with power spectral density (PSD) extracted as an energy-distribution feature. These diverse features form the dataset. A novel multi-class identification framework integrates a convolutional neural network (CNN) with t-SNE visualization, optimized via Adam to classify inner race faults, outer race defects, and rolling element failures. Experimental validation on a laboratory test rig with aligned training/testing loss curves confirms the proposed algorithm. The model achieves prediction accuracy over 98%, highlighting exceptional precision and generalization across fault types. This approach significantly advances bearing maintenance by developing noise-resistant algorithms that combines PSD energy features, RQA-derived recursive features, and reconstructed features, coupled with an integrated deep learning architecture to achieve state-of-the-art diagnostic reliability.