In order to make the best of abundant fault transient information, a neural network fault detecting method based on transient information is proposed. Firstly, a appropriate orthogonal wavelets packet function is chosen in order to effectively distill fault transient information; and then a chaotic neural network is employed to detect SPGFs, by which the interference of false and non- fault transient information is overcome, its weight coefficient and parameters are optimized by improved genetic algorithm; Finally, a numerical type of fault detecting criterion is designed. Effectiveness and advantage of the proposed method is tested by several experiments.