discusses the bp algorithm proposed for neural networks, which made the neural networks practical for use although the neural networks has not been applied properly because the user failed to stress the pre-treatment of input data neural networks and proposes an improved cauchy method of normalization of input data in the light of the characters of spacecraft, and methods of constructing normalization functions and concludes from comparison of these methods that normalization is the key reason having effect on the correctness of faults diagnosis in spacecraft bp算法的研制成功使神經(jīng)網(wǎng)絡達到了實用化程度.然而在實際工程中神經(jīng)網(wǎng)絡還沒有起到其應起的作用,主要原因不在神經(jīng)網(wǎng)絡本身而在各領域的使用者未能把重點放在輸入數(shù)據(jù)的前處理上.根據(jù)航天器故障診斷的特點,提出了根據(jù)數(shù)據(jù)特點來構造前處理函數(shù)的改進的升半柯西數(shù)據(jù)歸一化方法.通過比較證明,采用這一歸一化方法可大大提高神經(jīng)網(wǎng)絡故障診斷系統(tǒng)的準確性