How to add adaptive noise to model parameters?

The parameters of different layers have different orders of magnitude, if I add noise with the same sigma, the model will not convergence, I only have one globally calculated sigma, so how do I add noise adaptively depending on the magnitude of the data at each level, multiplication?