The numerous variables in non-steady state of aluminum reduction cells are non-Gaussian and impulsive.Due to correlation of variables

the condition that training samples must be independent and identically distributed is not fulfilled.For these reasons

it is too hard to diagnose the status of aluminum reduction cells in application.A diagnosis method for the status of aluminum reduction cells based on symmetric Alpha-stable(SαS)probabilistic distribution neural network was proposed.In the method

probability density function of SαS was introduced as radial basis function of model layer into the probabilistic neural network because such function had good fitting ability to non-Gaussian distributed data.And it also improved neural network approximation ability of partial pulse burst.By using 40 groups data of 170 kA operating aluminum smelter from a factory

this method could diagnose five statuses of aluminum reduction cells correctly and had not only stronger adaptability and robustness

but also approximation reliability and fast convergence.