Abstract: The majority of classification algorithm requires a large amount of labeled samples which are used to train the classifier model. In practical applications, samples which have been labeled are limited cause label them boring and expensive. The active learning and self-learning method based on uncertainty is introduced to the Support Vector Data Description（SVDD）, and a new classification method is proposed. Samples which have rich information to label them by experts are selecting through active learning, further, unlabeled sample set is used. The objective has good classification performance and saves the cost of labeling. The experimental results on identification of mechanical noise source of the submarine verify the performance of the proposed method.

Key words: active learning, self-learning, Support Vector Data Description（SVDD）, noise source identification

摘要： 多数分类识别算法需要大量的已标注样本对分类模型进行训练。实际应用中，对大量样本进行标注枯燥耗时且代价昂贵，因此能够获得的已标注样本数量非常有限。将基于不确定性样本的主动学习和代表性样本的自学习方法引入到基于支持向量数据描述的分类模型中，提出了一种新的分类识别方法。通过主动学习去挖掘那些对当前分类模型最有价值的样本进行人工标注，并借助自学习方法进一步利用样本集中大量的未标注样本，使得在花费较小的标注代价下，能够获得良好的分类性能。在潜艇机械噪声源识别问题上的实验结果验证了该方法能有效降低样本标注代价。

关键词: 主动学习, 自学习, 支持向量数据描述, 噪声源识别