Artificial-neural-network-based storage method for three-dimensional temperature field data during friction stir welding

In this paper, a new storage method for the three-dimensional temperature field data based on artificial neural network (ANN)was proposed. A multilayer perceptron that takes the coordinate (x,y,z) as inputs and temperature T as output, is used to fit the three-dimensional welding temperature field. Effect of number of ANN layers and number of neurons on the fitting errors is investigated. It is found that the errors decrease with the number of hidden layers and neural numbers per layers generally. When the number of hidden layers increases from 1 to 6, the maximum temperature error decreases from 74.74℃ to less than 2℃. The three-dimensional temperature field data is obtained by finite element simulation, and the experimental verification is completed by comparing the simulation peak temperatures with the measured results. As an example, an ANN with 4 hidden layers and 12 neurons in each layer were applied to test the performance of the proposed method in storage of the three-dimensional temperature field data during friction stir welding. It is found that the average error between the temperature data stored in ANN and the original simulation data that stored point-by-point is 0.517℃, and the error on the maximum temperature is 0.193℃, while the occupied disk space is only 0.27% of that is required in the conventional point-by-point storage.