Transfer Learning Technique for Automatic Bearing Fault Diagnosis in Induction Motors

In this paper, the transfer learning technique is used to modify two already trained Convolutional Neural Networks for bearing fault recognition. In particular, the transfer learning technique is used to transfer the acquired knowledge of the neural networks, a model of the AlexNet and "CommandNet", previously trained for image and speech recognition, to the particular case of the bearing fault recognition in induction motors. The transfer learning will substitute just the final layers of the Deep Neural Networks classifier, and requires for a complete re-training operation only few data in comparison to big training datasets required for a complete training from scratch. In this contribution vibration signals from two different sensors and datasets are used to generate the spectrograms to feed the networks.