在流式识别方法中,分块识别破坏并行性且消耗资源较大,而限制自注意力机制的上下文识别很难获得所有信息.由此,文中提出轻量化端到端声学架构(CFLASH-Transducer).为了获取细腻的局部特征,采用轻量化的FLASH(Fast Linear Attention with...在流式识别方法中,分块识别破坏并行性且消耗资源较大,而限制自注意力机制的上下文识别很难获得所有信息.由此,文中提出轻量化端到端声学架构(CFLASH-Transducer).为了获取细腻的局部特征,采用轻量化的FLASH(Fast Linear Attention with a Single Head)与卷积神经网络块结合.卷积块中采用Inception V2网络,提取语音信号多尺度的局部特征.再通过Coordinate Attention机制捕获特征的位置信息和多通道之间的相互关联.此外,采用深度可分离卷积,用于特征增强和层间平滑过渡.为了使其可流式化处理音频,采用RNN-T(Recurrent Neural Network Transducer)架构进行训练与解码.将当前块已经计算的全局注意力作为隐变量,传入后续块中,串联各块信息,保留训练的并行性和相关性,并且不会随着序列的增长而消耗计算资源.在开源数据集THCHS30上进行训练与测试,CFLASH-Transducer取得较高的识别率.并且相比离线识别,流式识别精度损失不超过1%.展开更多
The robust attitude control for a novel coaxial twelve-rotor UAV which has much greater payload capacity,higher drive capability and damage tolerance than a quad-rotor UAV is studied.Firstly,a dynamical and kinematica...The robust attitude control for a novel coaxial twelve-rotor UAV which has much greater payload capacity,higher drive capability and damage tolerance than a quad-rotor UAV is studied.Firstly,a dynamical and kinematical model for the coaxial twelve-rotor UAV is designed.Considering model uncertainties and external disturbances,a robust backstepping sliding mode control(BSMC)with self recurrent wavelet neural network(SRWNN)method is proposed as the attitude controller for the coaxial twelve-rotor.A combinative algorithm of backstepping control and sliding mode control has simplified design procedures with much stronger robustness benefiting from advantages of both controllers.SRWNN as the uncertainty observer is able to estimate the lumped uncertainties effectively.Then the uniformly ultimate stability of the twelve-rotor system is proved by Lyapunov stability theorem.Finally,the validity of the proposed robust control method adopted in the twelve-rotor UAV under model uncertainties and external disturbances are demonstrated via numerical simulations and twelve-rotor prototype experiments.展开更多
文摘在流式识别方法中,分块识别破坏并行性且消耗资源较大,而限制自注意力机制的上下文识别很难获得所有信息.由此,文中提出轻量化端到端声学架构(CFLASH-Transducer).为了获取细腻的局部特征,采用轻量化的FLASH(Fast Linear Attention with a Single Head)与卷积神经网络块结合.卷积块中采用Inception V2网络,提取语音信号多尺度的局部特征.再通过Coordinate Attention机制捕获特征的位置信息和多通道之间的相互关联.此外,采用深度可分离卷积,用于特征增强和层间平滑过渡.为了使其可流式化处理音频,采用RNN-T(Recurrent Neural Network Transducer)架构进行训练与解码.将当前块已经计算的全局注意力作为隐变量,传入后续块中,串联各块信息,保留训练的并行性和相关性,并且不会随着序列的增长而消耗计算资源.在开源数据集THCHS30上进行训练与测试,CFLASH-Transducer取得较高的识别率.并且相比离线识别,流式识别精度损失不超过1%.
基金Supported by the National Natural Science Foundation of China(No.11372309,61304017)Science and Technology Development Plan Key Project of Jilin Province(No.20150204074GX)the Science and Technology Special Fund Project of Provincial Academy Cooperation(No.2017SYHZ00024)
文摘The robust attitude control for a novel coaxial twelve-rotor UAV which has much greater payload capacity,higher drive capability and damage tolerance than a quad-rotor UAV is studied.Firstly,a dynamical and kinematical model for the coaxial twelve-rotor UAV is designed.Considering model uncertainties and external disturbances,a robust backstepping sliding mode control(BSMC)with self recurrent wavelet neural network(SRWNN)method is proposed as the attitude controller for the coaxial twelve-rotor.A combinative algorithm of backstepping control and sliding mode control has simplified design procedures with much stronger robustness benefiting from advantages of both controllers.SRWNN as the uncertainty observer is able to estimate the lumped uncertainties effectively.Then the uniformly ultimate stability of the twelve-rotor system is proved by Lyapunov stability theorem.Finally,the validity of the proposed robust control method adopted in the twelve-rotor UAV under model uncertainties and external disturbances are demonstrated via numerical simulations and twelve-rotor prototype experiments.
