一种基于TCM-SVDD的样本类别标注方法

A Method for the Identification of Sample Classification Based on TCM-SVDD

船舶机械噪声源的识别是一个小样本条件下的模式识别问题,采用增量学习是解决此问题的一条有效途径。但在进行增量学习以前,必须对新增样本的类别进行有效识别。为有效识别新增样本的类别,提出一种新的TCM-SVDD方法。首先,通过支持向量数据描述(SVDD)方法获得训练样本与新增样本的拉格朗日乘子;然后,将其作为该样本的奇异值代入直推置信机(TCM)中,估计新增样本属于不同类别的置信度,并将其与预设的置信水平进行比较;最后,剔除新增样本中的异类样本,实现增量学习。试验结果表明,该方法能快速、准确地识别异类模式样本,对训练样本集中混有少量异类模式样本的情况不敏感,而且可以控制对异类样本的检测准确率,自动化程度高。

The identification of the mechanical noise sources of a ship may be considered as a pattern rec-ognition problem based on small samples, which can be effectively solved by applying the increment learn-ing method. However, the new increment samples must be classified prior to the increment learning. Aim-ing at the problem, an innovative approach, named TCM-SVDD （Transductive Confidence Machine for Sup-port Vector Data Description）, is proposed in this paper for the classification of new samples. The strange-ness of all samples are first calculated using the SVDD method. Then, the confidence degree of new sam-ples are estimated and compared with the presettable confidence level. Finally, the identification of hetero-geneous pattern samples is accomplished. The results show that the heterogeneous pattern samples can be rapidly identified with the proposed method, and it is applicable even when there are only few heteroge-neous samples in the training sample set.