基于动态权重模型的数据不平衡SEI方法

Data imbalance SEI method based on dynamic weight model

摘要:

针对辐射源个体识别(SEI)中个体数据分布不平衡导致的识别准确率下降的问题，提出一种基于动态权重模型的SEI方法。通过设计一个动态类权重(DCW)模型，先利用元学习算法使用少量样本数据通过2层计算得到一个适中的权重初始值；再设计一种新的代价敏感损失函数计算预测值与真实值之间的距离，反向调整赋予少数类的学习权重，适度增加对少数类数据的重视程度。对少数类更加友好，对高度不平衡数据的处理有明显优势，缓解多数类样本对整个识别过程的计算误导，提高整体的识别正确率。

Abstract:

To tackle with the problem of decreased recognition accuracy caused by imbalanced individual data distribution in Specific Emitter Identification(SEI), a dynamic weight model based methodis proposed for individual identification of radiation sources. A Dynamic Class Weight(DCW) model is built. A moderate initial weight value is obtained by using a meta learning algorithm through two-layer calculation with a small amount of sample data. Then, a new cost sensitive loss function is designed to calculate the backward adjustment of the distance between the predicted value and the true value, which gives the minority learning weight, and moderately increases the attention to the minority data. It is more friendly to the minority. It has obvious advantages in the processing of highly unbalanced data, which alleviates the calculation misleading of the majority of samples in the whole recognition process, thus improving the overall recognition accuracy.