Damage detection and isolation from limited experimental data using simple simulations and knowledge transfer

A simulation model can provide insight into the characteristic behaviors of different health states of an actual system; however, such a simulation cannot account for all complexities in the system. This work proposes a transfer learning strategy that employs simple computer simulations for fault diagnosis in an actual system. A simple shaft-disk system was used to generate a substantial set of source data for three health states of a rotor system, and that data was used to train, validate, and test a customized deep neural network. The deep learning model, pretrained on simulation data, was used as a domain and class invariant generalized feature extractor, and the extracted features were processed with traditional machine learning algorithms. The experimental data sets of an RK4 rotor kit and a machinery fault simulator (MFS) were employed to assess the effectiveness of the proposed approach. The proposed method was also validated by comparing its performance with the pre-existing deep learning models of GoogleNet, VGG16, ResNet18, AlexNet, and SqueezeNet in terms of feature extraction, generalizability, computational cost, and size and parameters of the networks.