Identification of mechanical noise sources of underwater vehicle can be regarded as a pattern recognition problem using small sample. To solve the prob lem, an algorithm integrating the transductive confidence machine (TCM) with the Knearest neighbors (KNN) algorithm was usually employed, which is named the T C MKNN algorithm. But its classification ability is not satisfactory at the high confidence level. Therefore, an improved TCMIKNN algorithm is put forward by i m proving singular value measure method. The results of an experiment on a cabin m odel show that the improved TCMKNN algorithm can increase the classification a c curacy as well as the confidence level, and is superior to usual pattern recogni tion methods such as the BP and RBF artificial neural networks.