基于迁移学习的实时多类别文本分类方法研究

在实时文本分类任务中经常会有新类别出现,传统的文本分类方法通常难以利用实时的小规模样本学习新类别的模型。提出了一种基于迁移学习的新类别模型训练方法。首先将迁移学习技术和LS －SVM 算法结合,设计正则化项以控制源模型的迁移量及新模型与源模型的相似度,达到增量迁移。其次具体讨论了正则化项的学习方法。实验结果显示该方法相对同类算法有更好的分类性能,可以更高效地完成各种实时文本分类任务。     

高功率密度液压冲击器输出性能参数的确定

在介绍应用于液压驱动的非开挖地下穿孔机的主要部分－高功率密度液压冲击器的工作原理基础上，介绍了高功率密度液压冲击器输出性能参数确定。     

Dynamic compensator design and H∞ admissibilization for delayed singular jump systems via Moore–Penrose generalized inversion technique

This paper is intended to study output feedback-based $H_{\infty }$ admissibilization for singular Markovian jump time-varying delays systems via Moore–Penrose generalized inversion technique. The main aim is to design singular dynamic compensator based on output feedback control idea and Moore–Penrose generalized inversion technique, such that the time-varying delays singular Markovian jump system realizes not only stochastic stability, but also regularity, impulse-freeness (which are collectively known as stochastic admissibility) and achieves $H_{\infty }$ interference level. Time delay-dependent and mode-dependent L–K candidate functional is constructed, Moore–Penrose generalized inversion technique is utilized, then the improved $H_{\infty }$ admissibilization conditions for singular Markovian jump time-varying delays systems are provided by virtue of linear matrix inequalities. Simulation examples including a direct current motor-controlled inverted pendulum system are employed to confirm the effectiveness and practicability of the addressed singular dynamic compensation technique.