Abstract
Uniform temperature distribution during quenching thermal treatment is crucial for achieving exceptional mechanical and physical properties of alloy materials. Accurate and rapid prediction of the 3D transient temperature field model of large-scale aluminum alloy workpieces is key to realizing effective thermal treatment. This paper establishes a 3D transient temperature field model of large aluminum alloy workpieces and proposes a multi-loss consistency optimization-based physics-informed neural network (MCO-PINN) to realize soft sensing of the 3D temperature field model. The method is based on a MLP structure and adopts Gaussian activation functions. A surrogate model of the partial differential equation (PDE) is first constructed, and the residuals of the PDE, initial and boundary conditions, and observed data are encoded into the loss functions of the network. By establishing a Gaussian probability distribution model of each loss function and combining it with maximum likelihood estimation, the weight consistency optimization method of each loss function is then proposed to further improve the approximation ability of the model. To optimize the training speed of the network, an adaptive initial-value-eigenvector coding clustering (AIV-ECC) algorithm is finally proposed, which quickly determines the parameters of the Gaussian activation function, reduces the dependence on the initial value and improves the generalization performance of the network. Simulation and industrial experiments demonstrate that the proposed MCO-PINN can solve the 3D transient temperature field model with high precision and high time efficiency based on sparse measurements.