This work is focused on the realization of a low-complexity MobileNet neural network able to classify different bearing failure vibration signals. The network was designed to be deployed on an embedded microcontroller to be used in smart sensors in an IoT context, pursuing the paradigms of Embedded Artificial Intelligence and Edge Computing. Following these same concepts, data preprocessing as well has been kept as simple as possible, implementing a computationally low-cost strategy, allowing nonetheless for creating meaningful images (i.e. able to highlight the different bearing faults). Moreover, the network was trained with an ad hoc dataset created exploiting a test bench able to emulate different typologies of bearing failure vibration patterns, measured with different accelerometers. The study demonstrates, exploiting emulated data, how the MobileNet can generalize the learned features, exhibiting satisfactory performance when tested on data having different characteristics in terms of frequency components and acquired by sensors with different metrological behaviors, reaching an over-99% classification accuracy.