Recently, structural health monitoring has attracted considerable attention due to its great significance to prolong the lifespan and improve the system reliability of industrial machineries. The conventional fault diagnosis methods mainly depend on battery-support sensors. The self-powered fault diagnosis methods provide a new idea to realize fault diagnosis without using battery-supported or wired sensors. So, this study proposes a novel piezoelectric smart bearing for generating the voltage signals from the rotor’s vibrations. These voltage signals can be used for supplying the monitoring units and fault diagnosis. For this purpose, a prototype is fabricated and is experimentally tested with different shafts. Each shaft has a central disc and a breathing transverse crack at its mid-span. The shafts are identical except for the crack depth. Then, the convolutional neural network (CNN) algorithm is developed and trained based on the output voltage of the smart bearing under different crack depths. The results indicate that the crack fault can be classified to three labeled categories (level I, level II and level III). Also, the classification accuracy of this method can achieve 99.9%, which are quite feasible in practical applications. So, the proposed piezoelectric smart bearing has good application prospects for the self-powered fault diagnosis of rotating machinery.