Research on the Prediction Model of Milling Force and Temperature for High-Speed Milling of NAK80 Die Steel by Ultrasonic Atomization Micro-Lubrication Combined With Uniform Method and Back Propagation Neural Network

In this study, the high-speed milling of NAK80 die steel was carried out using the nanofluid (graphene) ultrasonic atomization micro-lubrication system combined with the uniform method and back propagation neural network. This study investigates the effects of rotational speed, feed, and nano-concentration on milling force and temperature and uses back propagation neural network (BPNN) to build an accurate prediction model. The uniform experimental design was first adopted. Utilizing the uniform dispersion feature reduces the time cost by using a smaller number of experimental groups. Use form $\mathrm{U}_{9}(9^{3})$ to design experiments and collect empirical data. Control factors include spindle speed, feed rate, and nanofluid concentration. Then, BPNN is used to establish the prediction model of milling force and temperature in high-speed milling process. To improve the prediction accuracy, the experiments will compare the effect of different hyperparameter settings on the accuracy of the BPNN model. Controlling and comparing hyperparameters includes using four other training functions: Levenberg-Marquardt, Trainglm, Traingda, Trainrp, and Trainscg, comparing the configuration of 5 and 10 neurons with different numbers of neurons to compare the accuracy and error value of the prediction model. The experimental results of the established model show that using the original nine sets of experimental data, the optimal hyperparameter settings of the milling force and milling temperature model are set to use Trainlm, the number of neurons is 10, and the MAPE value is 1.68%. To confirm the accuracy of the model, three different sets of experimental data were used and were not verified in the original parameter combination experiment. This verification part of the pre-model measured the best hyperparameters of the milling force and milling temperature model to use Trainlm. The number of neurons was 10, and the MAPE value was 4.45%. The results of this study show that a predictive model with high accuracy can be constructed using less experimental data, which can provide the industry with the ability to evaluate and adjust process parameters promptly to improve production efficiency when conducting related processes.