一种改进的LSTM模型在图像标题生成中的应用

为解决传统长短时记忆(LSTM)神经网络存在过早饱和的问题,使得对给定的图片能够生成更准确的描述,提出一种基于反正切函数的长短时记忆(ITLSTM)神经网络模型。首先,利用经典的卷积神经网络模型提取图像特征;然后,利用ITLSTM神经网络模型来表征图像对应的描述;最后在Flickr8K数据集上评估模型的性能,并与几种经典的图像标题生成模型如Google NIC等进行比较,实验结果表明本文提出的模型能够有效地提高图像标题生成的准确性。     

Backstepping approach for design of PID controller with guaranteed performance for micro-air UAV

Flight controllers for micro-air UAVs are generally designed using proportional-integral-derivative (PID) methods, where the tuning of gains is difficult and time-consuming, and performance is not guaranteed. In this paper, we develop a rigorous method based on the sliding mode analysis and nonlinear backstepping to design a PID controller with guaranteed performance. This technique provides the structure and gains for the PID controller, such that a robust and fast response of the UAV (unmanned aerial vehicle) for trajectory tracking is achieved. First, the second-order sliding variable errors are used in a rigorous nonlinear backstepping design to obtain guaranteed performance for the nonlinear UAV dynamics. Then, using a small angle approximation and rigorous geometric manipulations, this nonlinear design is converted into a PID controller whose structure is naturally determined through the backstepping procedure. PID gains that guarantee robust UAV performance are finally computed from the sliding mode gains and from stabilizing gains for tracking error dynamics. We prove that the desired Euler angles of the inner attitude controller loop are related to the dynamics of the outer backstepping tracker loop by inverse kinematics, which provides a seamless connection with existing built-in UAV attitude controllers. We implement the proposed method on actual UAV, and experimental flight tests prove the validity of these algorithms. It is seen that our PID design procedure yields tighter UAV performance than an existing popular PID control technique.     

基于LDA主题的改进TFIDF 95598工单智能分类研究

为了提高95595工单智能分类的准确率,提出了基于LDA(Latent Dirichlet Allocation)的改进TFIDF算法。先对文本提取特征词,然后采用K-means算法进行聚类处理。构建LDA模型,获得概率分布函数θ和φ,求取语义影响力SI(semantic influence,SI)作为特征词的权重,该改进算法记作SI-TFIDF(semantic influence-term frequency inverse document frequency,SI-TFIDF)。将SI-TFIDF算法与传统的TFIDF算法在sougou的数据库进行特征词提取,并采用K-means算进行聚类,对比结果显示,采用SI-TFIDF算法提取的特征词聚类效果优于TFIDF,验证了所提出方法的可靠性。在95598投诉工单上进行仿真实验,SI-TFIDF算法的投诉工单聚类准确率高于传统的TFIDF算法,验证了SI-TFIDF更适用于处理工单投诉的分类研究。     

Optimal robot‐environment interaction using inverse differential Riccati equation

An optimal robot‐environment interaction is designed by transforming an environment model into an optimal control problem. In the optimal control, the inverse differential Riccati equation is introduced as a fixed‐end‐point closed‐loop optimal control over a specific time interval. Then, the environment model, including interaction force, is formulated in a state equation, and the optimal trajectory is determined by minimizing a cost function. Position control is proposed, and the stability of the closed‐loop system is investigated using the Lyapunov direct method. Finally, theoretical developments are verified through numerical simulation.     

Improved recurrent neural networks for online solution of Moore‐Penrose inverse applied to redundant manipulator kinematic control

In this paper, two novel neural networks (NNNs), namely NNN‐L and NNN‐R neural models, are proposed to online left and right Moore‐Penrose inversion. As compared to GNN (gradient neural network) and the recently proposed ZNN (Zhang neural network) for the left or right Moore‐Penrose inverse solving, our models are theoretically proven to possess superior global convergence performance. More importantly, the proposed NNN‐R model is successfully applied to path‐tracking control of a three‐link planar robot manipulator. Illustrative examples well validate the theoretical analyses as well as demonstrate the feasibility of the proposed models, which are adopted and verified their effectiveness in kinematic control of a redundant manipulator, for real‐time Moore‐Penrose inverse solving.     

ISAR 系统相位失真的一种校正方法

(ISAR)主要应用于雷达静止不动而目标运动的场景,能够获得高分辨的二维图像。ISAR 系统比较 大的幅相失真会造成散焦等情况,从而使成像模糊,因此有必要对系统的幅相失真进行补偿以提高成像质量。文中给出某ISAR系统相位失真的一种实际校正办法。该系统带内相位原来800度左右的起伏,跟踪空间站时距离门内主副瓣比≤20 dB,经过该方法校正后带内相位达到≤5毅的水平,距离门内主副瓣比达到≤28 dB 的水平。     

基于时域延时的 ISAR 二维调制转发干扰

对逆合成孔径雷达(ISAR)来说,弹道中段目标的微多普勒效应实际上是一种方位向上的余弦调相效应,余弦调相 函数能实现频谱搬移,在方位向上产生多假目标。针对工程上无法实现对ISAR 信号方位向余弦调相的难题和方位向余 弦调相对距离向不起作用的问题,提出了一种基于时域延时的ISAR 二维调制转发干扰。该干扰在实现方位向余弦调相 的同时,能在距离向上形成一种新型盲移频的条带式干扰。 仿真结果表明:合适的延时参数将使目标淹没在距离-方位向上形成的多个假目标带中。     

深度逆向强化学习研究综述

深度逆向强化学习是机器学习领域的一个新的研究热点,它针对深度强化学习的回报函数难以获取问题,提出了通过专家示例轨迹重构回报函数的方法。首先介绍了3类深度强化学习方法的经典算法;接着阐述了经典的逆向强化学习算法,包括基于学徒学习、最大边际规划、结构化分类和概率模型形式化的方法;然后对深度逆向强化学习的一些前沿方向进行了综述,包括基于最大边际法的深度逆向强化学习、基于深度Q网络的深度逆向强化学习和基于最大熵模型的深度逆向强化学习和示例轨迹非专家情况下的逆向强化学习方法等。最后总结了深度逆向强化学习在算法、理论和应用方面存在的问题和发展方向。     

多传感器融合的数控机床热误差回归建模新方法

热误差对机床的加工精度影响很大,高性能的补偿系统依赖于多传感器融合建立的三维模型的精度、鲁棒性和合适的温度进行反馈输入。本文使用温度与位移传感器的模糊聚类进行温度分类,基于评价模型比对分析最优的温度分类,从每个分类中选择具有代表性温度作为候选温度。归纳试验数据,使用分段逆回归SIR模型进行热误差建模,SIR模型将高维前移回归问题转化为多个一维的回归问题,并且进一步消除了候选温度之间的耦合。热误差试验表明,SIR模型具有泛化能力强、预测精度高及鲁棒性好的特点,能够准确地描述多种典型工况条件下的实际热误差特性。     

基于分数阶傅里叶变换的ISAR横向定标

逆合成孔径雷达的横向定标是确定横向像素代表的真实尺寸、进行目标几何特征提取的前提。本文研究了一种基于分数阶傅里叶变换（Fractional Fourier transform,FrFT）的横向定标方法,利用回波信号在慢时间域为线性调频信号的特性,使用FrFT方法得到信号在不同分数阶傅里叶域的能量聚集性,估计目标回波信号多普勒调频率,求得转动角速度,进而完成横向定标。仿真实验验证了算法的有效性和准确性。