Convolution-weighting method for the physics-informed neural network: A Primal-Dual Optimization Perspective