Enhancing DP-SGD through Non-monotonous Adaptive Scaling Gradient Weight