基于联合域适应的异构样本增强网络

A Heterogeneous Sample Enhancement Network Based on Joint Domain Adaptation

域适应网络在样本增强领域应用受限,其根源在于领域的不同会加剧样本空间分布的差异。针对上述问题,提出基于联合域适应的异构样本增强网络。具体而言,异构域维度对齐子网中的支持域样本,在实现异构领域中样本维度对齐的同时,还嵌入了领域的分布知识,提高了后续异构域分布匹配的表现。此外,异构域分布匹配子网联合匹配了异构领域的边缘分布和条件分布,并嵌入了自适应机制,从而保证了联合域适应网络的匹配精度。由此,其他领域的样本通过上述设计的基于联合域适应的异构样本增强网络,能够被可靠地用于增强当前领域中的小样本。该网络在业界公认的田纳西-伊斯曼数据集上进行验证,实验结果表明了该网络的有效性。

The application of domain adaptation networks in the field of sample enhancement is limited,due to the fact that differences in domains lead to significant differences in the space distribution of samples.To address the above issues,a heterogeneous sample enhancement network based on joint domain adaptation is proposed.Specifically,the heterogeneous domain dimension alignment subnet,with designed support domain samples,achieves sample dimension alignment in heterogeneous domains while embedding the domain distribution knowledge,which helps to improve the performance of subsequent subnet.In addition,the heterogeneous domain distribution matching subnet,which jointly matches the marginal distribution and conditional distribution of the heterogeneous domains,embeds an adaptive mechanism,thus ensuring the matching accuracy.Furthermore,samples from other domains can be reliably used to enhance small samples through the network based on joint domain adaptation.The network is validated on the industry-recognized Tennessee-Eastman dataset,and the experimental results demonstrate the effectiveness of the network.