Title:Transfer Learning based Evolutionary Deep Neural Network for Intelligent Fault Diagnosis

Abstract:The performance of a deep neural network (DNN) for fault diagnosis is very much dependent on the network architecture. Also, the diagnostic performance is reduced if the model trained on a laboratory case machine is used on a test dataset from an industrial machine running under variable operating conditions. Thus there are two challenges for the intelligent fault diagnosis of industrial machines: (i) selection of suitable DNN architecture and (ii) domain adaptation for the change in operating conditions. Therefore, we propose an evolutionary Net2Net transformation (EvoNet2Net) that finds the best suitable DNN architecture for the given dataset. Nondominated sorting genetic algorithm II has been used to optimize the depth and width of the DNN architecture. We have formulated a transfer learning-based fitness evaluation scheme for faster evolution. It uses the concept of domain adaptation for quick learning of the data pattern in the target domain. Also, we have introduced a hybrid crossover technique for optimization of the depth and width of the deep neural network encoded in a chromosome. We have used the Case Western Reserve University dataset and Paderborn university dataset to demonstrate the effectiveness of the proposed framework for the selection of the best suitable architecture capable of excellent diagnostic performance, classification accuracy almost up to 100\%.