中文摘要：

辨识主要机械噪声源对于潜艇的噪声控制具有重要意义，但由于实际条件下测试的困难和昂贵的试验成本，通常难以获得足够多的训练样本，因此其本质上是一个小样本条件下的模式识别问题。为改善分类系统在小样本条件下的泛化性能，通过引入集成学习的BAGGING方法，分别与现有分类算法如分类与回归树（Classification and regressiontree，CART）和物差反传训练（Back-propagation，BP）相结合，提出了B-CART和B-BP算法。进一步，考虑到实际测量中往往同时利用布置在艇体不同部位上的多个通道（加速度传感器、水听器等）来采集数据，以期获得更多关于噪声源的相关信息，基于此先验信息提出了B-CART-M和B-BP-M算法。在此基础上，首先分别对每个通道的数据进行BAGGING集成，并生成该通道的结论，然后对每个通道的结论进行二次投票，从而得到最终分类结果，得到了算法B-CART-M’和B-BP-M’。舱段模型试验结果表明，以上6种算法均能不同程度提高小样本条件下分类系统的性能，其中B-CART-M’和B-BP-M’效果最为明显；对同一算法而言，外壳数据的分类效果最好，远场数据的分类效果最差，内壳和近场数据的分类效果相差不多。给出了算法实际应用时的若干建议。

英文摘要：

Identification of major mechanical noise sources is a key step of the noise control in submarine. But it is usually difficult to obtain enough training samples due to the high expense of testing. So it can be regarded as a pattern recognition problem on sparse data. In order to improve the generalization ability of the classifier, six different algorithms are proposed. The BAGGING method is chosen as the ensemble algorithm. And without the loss of generality, the classification and regression tree （CART） and back-propagation （BP） algorithm are chosen as basic classifiers. Simply combing BAGGING with CART and BP algorithm respectively, two algorithm called B-CART and B-BP algorithm are proposed. Considering the priori： ① more than one data channel （accelerometer, hydrophone, etc.） is frequently used in most cases; ② fusing the information from multiple sensors can enhance the reliability of the classification rate, two algorithms called B-CART-M and B-BP-M are proposed. Furthermore, after calculating the classification result of signal in one channel by BAGGING and plurality voting first, the final classification result from all sensory channels is obtained by the second voting, which is the main idea of algorithms called B-CART-M＇ and B-BP-M＇. Measurements on a submerged full scale model section of a submarine hull showed： ① all the six algorithms proposed above can improve the classification rate and the performance of B-CART-M＇ and B-BP-M＇ is the best; ② for the given algorithm, the classification rate of outer shell data is the highest and far field lowest. Several suggestions of application of the six proposed algorithms are given.