Application of combined transfer learning and convolutional neural networks to optimize plasma spraying

Abstract

Deep transfer learning can make full use of pre-trained neural networks and has been used in many cases with limited sample data. In this work, parameter-transfer learning was implemented to model the relationship between process control parameters and in-flight particle behavior. Six different parameter-transfer learning models were designed to fine-tune the variables of the convolutional neural network (CNN) model pre-trained with a dataset of yttria-stabilized zirconia (YSZ) particles. Then transfer learning models were trained with the new dataset obtained from simulation results of NiCrAlY particles and the losses of different models in the training set and test set were compared. Results indicate that the method in which the entire pre-trained CNN model was fine-tuned, combined with a decreasing learning rate, exhibited the lowest loss in the training dataset and the highest testing accuracy. Particle status distributions obtained from the control parameters predicted by the transfer learning model were found to be in good agreement with the corresponding designed target values.