Prediction of control parameters corresponding to in-flight particles in atmospheric plasma spray employing convolutional neural networks

Optimization of control parameters for plasma spraying process is of great importance in thermal spray technology development. Engineers may limit themselves to local optimal solution by considering countable potential design solutions when selecting the plasma spray parameters in practice. This work proposes one decision support model by employing convolutional neural network (CNN) to explore the suitability of preliminary design. The approach aims to help engineers select global optimal solution with short time and low labor cost by invoking the models' capability of extracting potential features of in-flight particle characteristics. Simulation results make it possible to analyze new spraying process and train the designed model under the condition of insufficient experience and data. Firstly, the distributions of particle status obtained from simulation results act as the input and the control parameters are the output. Secondly, the projections between the in-flight particles and the control parameters are built implicitly and analyzed through CNN models. Thirdly, we validate the statistical information of particle state distributions through visualizing the feature maps and filters. Finally, the trained CNN models are verified by the fitted Gaussian distributions with basically consistent results. By building projections of in-flight particles and control parameters, new entrants and apprentices are capable of deducing the control parameters with the help of the pre-trained CNN model, thus cutting down the threshold for new practitioners.