一种具有较强泛化能力的神经网络模型研究与应用

分析了影响神经网络模型泛化能力的因素。以电解铝过程中氧化铝浓度的神经网络软测量为例 ,提出了利用先验知识确定网络结构 ,采用特定实验保证样本数量和质量 ,离线训练加在线学习修正模型等措施 ,改善神经网络模型的泛化能力。现场应用表明这些措施是有效的。这样建立的神经网络模型准确 ,泛化能力强 ,为实现过程的先进控制提供了可靠保障     

基于神经网络的导弹姿态控制系统故障诊断

本文提出了一种多层前向神经网络快速误差向后传播学习算法ＦＢＰ（ＦａｓｔｅｒｒｏｒＢａｃｋＰｒｏｐａｇａｔｉｏｎＬｅａｒｎｉｎｇＡｌｇｏｒｉｔｈｍ），通过某导弹姿态控制系统故障诊断的仿真研究，验证了ＢＰ（误差后向传播学习算法）和ＦＢＰ用于导弹姿态控制系统故障诊断的有效性，ＦＢＰ学习算法较之ＢＰ学习算法学习速度的快速性     

再入航天器的神经网络自适应控制

对于再入航天器来说 ,可恢复故障通常会导致总体控制权限的降低。本文将伪控制限制方法和神经网络自适应控制运用到再入航天器的制导与控制中。利用两个控制回路来实现控制目标。其中 ,外回路适用于力的扰动 ,而内回路适用于力矩的扰动。限制外回路以阻止对内回路动态特性的调整。在达到控制极限时 ,伪控制限制也可进行自适应调整。另外 ,本文还通过加入加速度计反馈元件以提高系统的性能     

多模式组合诊断技术在导弹导引头上的应用研究

阐明了四种类型的多模式组合诊断方法及其故障诊断的核心思想。以某导弹导引头的故障诊断为实例，详细描述了基于模糊理论、神经网络和粗糙集理论的三组合模式的诊断过程，并给出了诊断结果；介绍了我们研制的基于不确定性理论和专家系统组合的智能化故障诊断专家系统的应用。研究和应用实例表明：多模式组合诊断技术通用性和鲁棒性强。能够加快诊断速度。降低误诊率。尤其适合于复杂系统的故障诊断，具有很大的工程实用价值。     

基于BP人工神经网络的GPS／SINS组合导航算法

基于扩展Kalman滤波的GPS／SINS组合导航算法，需要对原始的非线性连续系统模型进行线性化和离散化处理，要求系统噪声和测量噪声为零均值的高斯白噪声，且易于出现滤波器发散。BP人工神经网络毋需对所求解的问题建模，能够很好地逼近系统非线性特性，获得较高精度的导航定位信息；还具有计算过程稳定，不涉及矩阵求逆，不需要迭代逼近，以及容易实现并行处理等优点。本文设计适用于GPS／SINS组合导航系统的BP网络模型，并在标准的BP算法基础上，采用共轭梯度法改进网络训练速度及精度。最后，通过仿真算例说明BP网络方法用于GPS／SINS组合导航计算的可行性。     

An Artificial Neural Network based approach for estimation of rain intensity from spectral moments of a Doppler Weather Radar

By using a Doppler Weather Radar (DWR) at Shriharikota (13.66°N & 80.23°E), an Artificial Neural Network (ANN) based technique is proposed to improve the accuracy of rain intensity estimation. Three spectral moments of a Doppler spectra are utilized as an input data to an ANN. Rain intensity, as measured by the tipping bucket rain gauges around the DWR station, are considered as a target values for the given inputs. Rain intensity as estimated by the developed ANN model is validated by the rain gauges measurements. With the help of a developed technique, reasonable improvement in the estimation of rain intensity is observed. By using the developed technique, root mean square error and bias are reduced in the range of 34–18% and 17–3% respectively, compared to Z–R approach.     

惯性推算误差抑制的神经网络自学习模型设计与验证

惯性推算误差抑制是提升复杂场景下组合导航定位性能的关键,现有采用运动约束或系统误差高阶建模的方法从运动学模型及传感器误差模型出发,通过经验确定参数及模型的最优解。深度学习隐式模型能够挖掘数据之间的隐含关系,进行自主化参数寻优,并在提升惯导误差建模精度方面具有一定优势。总结了现有主流网络模型设计的优缺点,通过对比不同的输入输出方案进行优选,最终利用卷积神经网络构建了一套惯性误差抑制的轻量化神经网络自学习模型,并利用实测车载数据验证了该模型的有效性。实验结果表明,在GNSS信号失锁300 s的路段I和失锁285 s的路段II,网络模型速度约束的算法相较于纯惯性推算和传统NHC算法均有一定提升,融合NHC及网络模型速度约束的算法在水平定位精度上分别改善了41.7%~47.4%和26.7%~36.6%,一定程度上抑制了惯性推算误差。     

Numerical analysis and optimization of boundary layer suction on airfoils

Numerical approach of hybrid laminar flow control(HLFC) is investigated for the suction hole with a width between 0.5 mm and 7 mm. The accuracy of Menter and Langtry’s transition model applied for simulating the flow with boundary layer suction is validated. The experiment data are compared with the computational results. The solutions show that this transition model can predict the transition position with suction control accurately. A well designed laminar airfoil is selected in the present research. For suction control with a single hole, the physical mechanism of suction control, including the impact of suction coefficient and the width and position of the suction hole on control results, is analyzed. The single hole simulation results indicate that it is favorable for transition delay and drag reduction to increase the suction coefficient and set the hole position closer to the trailing edge properly. The modified radial basis function(RBF) neural network and the modified differential evolution algorithm are used to optimize the design for suction control with three holes. The design variables are suction coefficient, hole width, hole position and hole spacing. The optimization target is to obtain the minimum drag coefficient. After optimization,the transition delay can be up to 17% and the aerodynamic drag coefficient can decrease by 12.1%.     

A new global version of M(3000)F2 prediction model based on artificial neural networks

A new version of global empirical model for the ionospheric propagation factor, M(3000)F2 prediction is presented. Artificial neural network (ANN) technique was employed by considering the relevant geophysical input parameters which are known to influence the M(3000)F2 parameter. This new version is an update to the previous neural network based M(3000)F2 global model developed by Oyeyemi et al. (2007), and aims to address the inadequacy of the International Reference Ionosphere (IRI) M(3000)F2 model (the International Radio Consultative Committee (CCIR) M(3000)F2 model). The M(3000)F2 has been found to be relatively inaccurate in representing the diurnal structure of the low latitude region and the equatorial ionosphere. In particular, the existing hmF2 IRI model is unable to reproduce the sharp post-sunset drop in M(3000)F2 values, which correspond to a sharp post-sunset peak in the peak height of the F2 layer, hmF2. Data from 80 ionospheric stations globally, including a good number of stations in the low latitude region were considered for this work. M(3000)F2 hourly values from 1987 to 2008, spanning all periods of low and high solar activity were used for model development and verification process. The ability of the new model to predict the M(3000)F2 parameter especially in the low latitude and equatorial regions, which is known to be problematic for the existing IRI model is demonstrated.