Improving performance in pulse radar detection using neuralnetworks

A new approach using a multilayered feed forward neural network for pulse compression is presented. The 13 element Barker code was used as the signal code. In training this network, the extended Kalman filtering (EKF)-based learning algorithm which has faster convergence speed than the conventional backpropagation (BP) algorithm was used. This approach has yielded output peak signal to sidelobe ratios which are much superior to those obtained with the BP algorithm. Further, for use of this neural network for real time processing, parallel implementation of the EKF-based learning algorithm is indispensable. Therefore, parallel implementation has also been developed     

Intelligent modeling and identification of aircraft nonlinear flight

In this paper, a new approach has been proposed to identify and model the dynamics of a highly maneuverable fighter aircraft through artificial neural networks（ANNs）. In general, aircraft flight dynamics is considered as a nonlinear and coupled system whose modeling through ANNs, unlike classical approaches, does not require any aerodynamic or propulsion information and a few flight test data seem sufficient. In this study, for identification and modeling of the aircraft dynamics, two known structures of internal and external recurrent neural networks（RNNs） and a proposed structure called hybrid combined recurrent neural network have been used and compared.In order to improve the training process, an appropriate evolutionary method has been applied to simultaneously train and optimize the parameters of ANNs. In this research, it has been shown that six ANNs each with three inputs and one output, trained by flight test data, can model the dynamic behavior of the highly maneuverable aircraft with acceptable accuracy and without any priori knowledge about the system.     

基于反馈误差学习的神经网络控制

研究了应用神经网络和ＰＤ反馈控制实现非线性系统的自适应跟踪问题。ＰＤ反馈控制器不但保证闭环系统的稳定性,同时其输出又作为训练神网络的参考信号。证明了通过选择适当的初始加权和加权的调节速率可以实现非线性系统的固定点跟踪。     

基于小波变换的容差电路故障诊断

针对目前模拟电路中电子元器件存在的容差与非线性导致电路故障难以检测的现状,设计了适用于诊断由器件超出容差所引起的模拟电路故障的小波分析诊断方法。通过设定故障进行蒙特卡罗容差实验,采用小波神经网络,对故障输出信号进行小波分析提取其小波高频系数参量,经PCA分析和归一化后形成训练特征向量,并经过BP神经网络训练后,故障信号通过小波神经网络后能够快速精准的对故障器件进行定位。通过大量样本进行仿真计算表明所设计的小波特征参量故障诊断法对于模拟电路具有很好的故障分辨率。     

试验设计法与BP-GA算法在复合材料结构优化中的应用

试验设计法具有均衡分散性和整齐可比性,通过试验设计进行响应面的样本选取可以大大减少样本的数目,同时保证响应面的训练精度。响应面与遗传算法相结合在工程应用中已经形成一套适用于复杂结构设计的高效优化方法。本文以复合材料加筋蒙皮的质量优化问题为例,通过正交试验进行响应面训练样本选取,通过PATRAN/NASTRAN进行有限元分析得出样本的数值,应用神经网络(BP)训练、构建优化的目标函数和强度、稳定性约束响应面,结合其他常规约束条件通过遗传算法(GA)进行优化,验证了这套算法的有效性和实用性。     

液体火箭发动机高速涡轮泵的振动故障检测

讨论了涡轮泵故障的几个主要原因，据此提取涡轮泵振动数据的特征，用BP神经网络的方法进行故障检测。BP神经网络的训练样本集由一个具有无监督聚类功能的神经网络从原始的特征向量集获取。     

基于RBFNN自适应混合学习算法的航天测控系统任务可靠性分配

为解决执行航天测控任务的各设备存在复杂的时空关联、可视与信息关联等动态约束关系,使得航天测控系统任务可靠性分配建模和分析极其困难,同时模型求解效率低的问题,提出了自适应混合学习算法的径向基神经网络建模方法.算法通过训练样本相关性矩阵的主成分分析确定网络隐含层初始节点数;在此基础上,利用梯度信息衰减因子改进了迭代过程中网络参数的梯度信息计算方式,避免了学习过程早熟的不足,且加快了迭代收敛速度.最后,通过采集航天测控系统输入-输出数据,将自适应混合学习算法应用于参数训练,并给出了具体实现步骤.通过算例仿真,表明算法在解决航天测控系统任务可靠性分配问题时具有较高泛化能力和分配结果稳定等优点.      

神经网络在发动机自适应建模中的应用研究

提出了一种新的基于神经网络的发动机自适应实时模型的建模方法。建模的思想是认为发动机的任何非额定工作都将导致其输出参数的变化,因而可以把这些参数偏离正常工作参数值的变化量,也就是输出偏离量,用来表征发动机的非额定工作情况。把它们作为增广的状态变量,设计卡尔曼滤波器对其进行最优估计,然后用这些输出偏离量的估计值,通过由BP神经网络训练出来的可测输出偏离量与未测输出偏离量的映射关系来校正机载发动机模型的计算输出,使之与真实发动机的输出一致,从而使实时机载模型获得对任何发动机非额定工况的自适应能力。      

脉冲风洞一体化飞行器测力精度分析

为评估脉冲风洞一体化飞行器的测力精度,根据脉冲风洞测力系统特点,对其进行简化,建立其动力学方程,并获得其动态脉冲响应衰减函数;分析测力系统输出信号变化规律,对其进行傅里叶变换,获得系统振动主频,并运用神经网络、梯度下降法对天平信号进行准确拟合,预测了测力信号的趋势值;对多次试验结果进行了信号拟合和趋势预测,获得了天平测力的精度。结果表明:该方法能够准确预测天平稳定输出结果,且当天平输出信号达到4个周期时,测力精度能够达到97%以上。      

一个神经网络结合遗传算法的叶轮逆命题设计方法

本文的目的是利用神经网络实现离心压缩机叶轮的逆命题设计。为此,结合实例叶轮逆命题设计问题通过遗传算法的演化和训练建立了前馈神经网络。所建网络随之用于求解给定表面速度分布情况下的叶片型线。      

航空发动机神经网络自学习PID控制

将神经网络与传统的PID控制相结合,构成神经网络自学习PID控制,用神经网络在线整定PID控制器的比例、积分及微分三个参数,使被控对象跟踪理想参考模型的输出。该系统具有自学习能力,能适用于非线性、时变的被控对象。将神经网络自学习PID控制方法用于航空发动机全包线控制以及蜕化发动机的控制,进行了数字仿真,验证了该方法的有效性。     

航空发动机递归神经网络分路式解耦控制

针对航空发动机多变量控制中变量之间的耦合问题,提出了一种基于递归神经网络的分路式动态解耦控制方法,给出了发动机双路式解耦控制系统的结构及其解耦原理和算法。利用递归小波网络较强的动态非线性映射能力,在线完成发动机各控制通道的模型辨识,并回馈对应的灵敏度信息;神经网络PID控制器根据回馈的信息在线自适应调整参数,实现发动机各通道的准确跟踪和分路独立控制。仿真表明,该方法在保证控制系统良好的动态和稳态性能的同时,有效地减小了各回路之间的耦合影响,能够成功应用于发动机控制系统的解耦。      

基于神经网络的鲁棒制导律设计

 基于神经网络理论对寻的导弹鲁棒制导律进行了优化设计。建立了制导系统非线性运动学方程和鲁棒性能函数,并将鲁棒性能函数转化成了微分对策的极小极大化问题。采用伴随 BP技术,将微分对策的两点边值求解问题转化为 2个神经网络的学习问题,训练后的 2个神经网络分别作为对策双方的最优控制器在线使用,避免了直接求解复杂的鲁棒制导律问题,仿真结果表明了该方法有效性。     

Signal Resolution via Digital Inverse Filtering

Research in numerous areas is directed toward the resolution of multiple overlapping signals in a noisy environment. These areas include radar, sonar, speech, seismology, and electrophysiology. Sometimes matched filters are used; other times inverse filters are employed. This paper discusses one approach to the analysis of the resolution of inverse filters. Our method is to compromise the trade-off between signal resolution and the output signal-to-noise ratio (SNR). A performance measure for the inverse or deconvolution filter is defined as a quantity proportional to the harmonic mean of the resolution and the SNR. An optimum output pulse duration is obtained using this criterion, where the pulse shape has been previously selected and the input signal waveform is known. In addition, upper and lower bounds for the output pulse duration are presented. Graphs are given which allow the designer to select the optimum inverse filter output pulse duration for a desired signal resolution and an estimated SNR.     

An intelligentized and fast calibration method of SINS on moving base for planed missiles

In order to minish the error of inertial sensors, the technology of neural networks is attempted to on-line calibration of a slave inertial navigation system mounted on planed missiles. Based on the time-varied specialty of slave inertial navigation system on a moving base, an input–output sample structure method is proposed, and to automatically calibrate and revise the error of inertial sensors of inertial navigation system. When a missile is appended under the wing and in free-flight, in order to solve the inconsistent problem of measurement's character of the inertial sensors, the error angles between the master inertial navigation system and the slave inertial navigation system are estimated in advance, then, the input samples of a neural network can correctly simulate the free-flight state. Furthermore, in order to make a learning algorithm of neural networks can satisfy real-time calibrating on a moving base, the traditional Newton algorithm is improved by using first differential coefficient to replace the approximate matrix of second differential coefficients. As a result, the training speed and precision of neural network are enhanced. The simulation results indicate that the method and algorithm are feasible.     

Neural network modeling of flight control system

Modeling of angle tracking systems in the presence of actuator non-linearity such as angle, position and rate limits is a very significant and difficult task in the design and implementation of aircraft, target-tracking, and missile guided systems. A new recurrent neural network with time-delayed inputs and output feedback is used for the modeling of angle tracking systems, with emphasis on the neural network architecture, principles and algorithms. The neural network controller with modeling units for angle tracking is designed by using TMS320C25 processors. For time and size requirements, limited precision technology and look-up table technology are used in the design of the hardware and software systems. Given a set of input commands, the network is trained to control the system within the constraints imposed by actuators. The results show that the proposed networks are able to model the angle tracking system through learning without separate consideration of the non-linearity of actuators     

Online INS/GPS integration with a radial basis function neural network

Most of the present navigation systems rely on Kalman filtering to fuse data from global positioning system (GPS) and the inertial navigation system (INS). In general, INS/GPS integration provides reliable navigation solutions by overcoming each of their shortcomings, including signal blockage for GPS and growth of position errors with time for INS. Present Kalman filtering INS/GPS integration techniques have some inadequacies related to the stochastic error models of inertial sensors, immunity to noise, and observability. This paper aims to introduce a multi-sensor system integration approach for fusing data from INS and GPS utilizing artificial neural networks (ANN). A multi-layer perceptron ANN has been recently suggested to fuse data from INS and differential GPS (DGPS). Although being able to improve the positioning accuracy, the complexity associated with both the architecture of multi-layer perceptron networks and its online training algorithms limit the real-time capabilities of this technique. This article, therefore, suggests the use of an alternative ANN architecture. This architecture is based on radial basis function (RBF) neural networks, which generally have simpler architecture and faster training procedures than multi-layer perceptron networks. The INS and GPS data are first processed using wavelet multi-resolution analysis (WRMA) before being applied to the RBF network. The WMRA is used to compare the INS and GPS position outputs at different resolution levels. The RBF-ANN module is then trained to predict the INS position errors and provide accurate positioning of the moving platform. Field-test results have demonstrated that substantial improvement in INS/GPS positioning accuracy could be obtained by applying the combined WRMA and RBF-ANN modules.     

Classification of radar targets using synthetic neural networks

Radar target classification performance of neural networks is evaluated. Time-domain and frequency-domain target features are considered. The sensitivity of the neural network algorithm to changes in network topology and training noise level is examined. The problem of classifying radar targets at unknown aspect angles is considered. The performance of the neural network algorithms is compared with that of decision-theoretic classifiers. Neural networks can be effectively used as radar target classification algorithms with an expected performance within 10 dB (worst case) of the optimum classifier     

一种小波神经网络与遗传算法结合的优化方法

提出一种基于小波神经网络(简称WNN)与Pareto遗传算法相结合的优化方法,并用于内流的数值流场优化计算.小波神经网络由输入层、隐含层和输出层组成.在隐含层用Morlet小波母函数取代了误差反向传播(BP)神经网络中常用的Sigmoid激励函数.Pareto遗传算法具有很好的全局寻优能力和良好的优化效率,在通常情况下它总可以得到均匀分布的Pareto最优解集.典型算例表明:该算法快速、高效,能高精度的完成非线性函数的逼近与映射,其泛化能力很强.      

基于卷积门控循环网络的滚动轴承故障诊断

针对许多基于深度学习的滚动轴承故障诊断方法在小样本数据集下诊断性能下降的问题,提出一种基于卷积门控循环神经网络的轴承故障诊断模型。该模型使用两层的卷积网络来从输入信号中提取特征,同时使用tanh函数作为激活函数,且池化层使用大池化核来进行重叠下采样。将所提取得到的高层特征连接到双向门控循环网络。合并循环网络正向和逆向的最后一个状态,并连接一层全连接层进行输出。选用凯斯西储大学的轴承故障数据集来验证模型在小样本数据集下的诊断性能,实验结果表明,相比于其他类型的模型,该模型在仅有20个训练样本的情况下依然保持97%的识别准确率。