无人机类脑吸引子神经网络导航技术

当前无人机在非结构化或未知环境下飞行主要采用SLAM进行导航与定位,存在如下突出问题:依赖高精度昂贵激光雷达等环境感知传感器;需要建立准确世界和无人机物理模型;受环境影响较大;自主智能水平较低,无法较好地满足无人机对导航系统的要求,需要发展自主智能的导航方式。基于吸引子神经网络的类脑导航技术,无需训练模型参数,不依赖高精度传感器,无需精确建模,且复杂环境下鲁棒性较强,具有解决上述问题的潜力。简要阐述了动物大脑导航机理,分析了吸引子神经网络和基于吸引子神经网络的类脑导航关键技术,最后讨论了吸引子类脑导航技术在无人机应用中的挑战。

基于神经网络模型分割的入侵检测方法

针对神经网络在入侵检测应用中存在资源消耗大、学习效率低等不足,提出一种基于神经网络模型分割的入侵检测方法。该方法根据当前典型攻击的特征,为每类攻击分别建立独立的子神经网络,对该类攻击进行学习和检测。然后再将每个子神经网络分割成多个更小的子模型,来降低学习时间和减少神经网络各层之间的连接权数目。设计了相应算法并进行仿真实验。实验结果表明,提出的方法提高了入侵检测的速度,降低了系统资源的消耗,提高了检测率。

基于子块集成神经网络法的PSD背景光补偿

提出一种基于神经网络高精度线性化子块网络集成插值实现光电位置敏感器件(PSD)背景光非线性补偿方法。利用神经网络具有逼近任意非线性函数的特点,通过训练,使神经网络建立在不同背景光下PSD输出与其标准值之间的非线性映射关系,实现PSD全程跟踪补偿。实验结果表明,该方法能有效地消除背景光的影响,在神经网络的输出端得到期望的线性输出。

未知复杂环境下基于兴趣驱动的类脑自主导航技术

随着无人系统的应用越发广泛,传统导航技术很难满足无人系统在面对复杂任务和未知环境时对自主智能导航性能的要求。哺乳动物能够在兴趣驱动下实现高效智能且自适应环境的导航行为,受此启发的基于兴趣驱动的类脑自主导航技术有潜力克服传统导航实时、准确和低功耗不能同时满足的缺点。本文首先阐述了哺乳动物大脑导航机理;其次,总结概括出基于兴趣驱动的类脑自主导航技术框架;再次,从自身感知、环境认知、记忆推理和兴趣决策4个方面梳理了类脑自主导航的关键技术和实现途径,指出了相关研究的缺陷;最后,分析了现阶段类脑自主导航技术的不足,并对未来一体化发展做出展望。

基于神经网络混沌吸引子的混合加密

把Diffe-Hellman密钥交换协议和流密码算法相结合,设计了一种基于神经网络混沌吸引子的混合加密算法。算法采用基于混沌吸引子的Diffe-Hellman公钥体制,保证了密钥分发的安全性,同时拥有流密码速度快的优点,提高了加密速度,因此实用性较好,能够满足下一代通信实时快速的需求。分析了算法的安全性和加解密效率,利用vc编程实现算法,并对仿真生成的密钥流和密文进行测试。实验结果表明,算法具有较好的安全性和加解密速度。

运动目标预测跟踪的神经计算机制

迄今为止,人们对神经系统如何提取外部输入中静态信息的机制已经有所了解,但对其处理动态信息的机制却知之甚少.在处理运动信息时,神经系统面临的一个根本性挑战是克服神经信号在大脑内传输的时间延迟.这种延迟是显著的,同时又是不可避免的.它是层次化的神经信号通路和模块化的脑功能分区在传递、交流信息时必然产生的结果.如果这些时间延迟得不到补偿,神经系统对快速运动物体的空间位置的感知就会滞后于物体的真实位置,从而不能实时处理运动信息.大量实验表明大脑补偿延迟的策略是对运动物体将要到达的空间位置做出预测.回顾已有的数理模型,主要是具有负反馈效应的连续吸引子神经网络模型,及其实现运动预测的神经计算机制,并对相关神经网络模型在类脑计算中的应用前景进行了展望.

基于深度网络分级特征图的图像超分辨率重建

从低分辨率图像中提取特征图恢复高分辨率图像中的高频信息是超分辨率重建的一个关键问题,针对该问题提出一个新的基于卷积神经网络的超分辨率重建算法.网络结构由卷积层与子像素卷积组成,特征提取网络中卷积层提取低分辨率图像的特征,重建网络中子像素卷积神经网络作为上采样算子.针对不能充分利用多级特征图的问题,采用跳跃连接和特征图联结在特征提取网络末端跨通道融合特征图,同时降低特征图的维度.并在此基础上再次提取特征图应用于重建.实验结果表明,算法在PSNR、SSIM和人类视觉效果上与其他基于深度学习的算法相比有着显著的提高.

基于频谱及轴心轨迹图的汽轮机故障诊断

采用频谱及轴心轨迹图的方法提取仿真台得到的故障振动信号特征,分别建立子BP神经网络,并采用D-S证据理论对子BP神经网络的输出进行融合(多层信息融合)方法,从不同侧面对故障进行诊断。结果表明:采用多层信息融合方法的故障诊断置信度比频谱方法提高约0.03,比轴心轨迹图方法提高0.4,效果显著;对故障类型的识别准确率具有显著提高。

基于双目视觉的类脑三维认知地图构建方法

基于大脑导航神经细胞机理的类脑认知地图构建方法,为发展智能同步定位与地图构建(SLAM)技术提供了新思路。针对现有类脑认知地图构建精度不高的问题,提出了一种基于双目视觉的类脑三维认知地图构建方法。首先阐述了类脑三维认知地图系统的工作原理,然后论述了不同视觉里程计对认知地图精度的影响,研究了基于双目视觉里程计的类脑三维认知地图精度优化方法,最终完成了基于视觉数据集的类脑三维认知地图构建试验。试验结果表明,所提方法构建的视觉里程计地图的三维位置误差为总行程的2.14%,认知地图的三维位置误差为1.56%;认知地图精度与里程计精度呈正相关;系统通过模板匹配进行回环检测与校正,提高了认知地图的精度。

Research on Thermodynamic Properties of Polybrominated Diphenylamine by Neural Network

Based on the location of bromine substituents and conjugation matrix, a new substituent po- sition index ~X not only was defined, but also molecular shape indexes Km and electronega- tivity distance vectors Mm of diphenylamine and 209 kinds of polybrominated diphenylamine （PBDPA） molecules were calculated. Then the quantitative structure-property relationships （QSPR） among the thermodynamic properties of 210 organic pollutants and 0X, K3, M29, M36 were founded by Leaps-and-Bounds regression. Using the four structural parameters as input neurons of the artificial neural network, three satisfactory QSPR models with network structures of 4：21：1, 4：24：1, and 4：24：1 respectively, were achieved by the back-propagation algorithm. The total correlation coefficients R were 0.9999, 0.9997, and 0.9995 respectively and the standard errors S were 1.036, 1.469, and 1.510 respectively. The relative mean deviation between the predicted value and the experimental value of Sθ, AfHe and △fGθ- were 0.11%, 0.34% and 0.24% respectively, which indicated that the QSPR models had good stability and superior predictive ability. The results showed that there were good nonlinear correlations between the thermodynamic properties of PBDPAs and the four structural pa- rameters. Thus, it was concluded that the ANN models established by the new substituent position index were fully applicable to predict properties of PBDPAs.

Construction of Early-warning Model for Plant Diseases and Pests Based on Improved Neural Network

By studying principles and methods related to early-warning model of plant diseases and using PSO method, parameter optimization was conducted to backward propagation neural network, and a pre-warning model for plant diseases based on particle swarm and neural network algorithm was established. The test results showed that the construction of early-warning model is effective and feasible, which will provide a via- ble model structure to establish the effective early-warning platform.

Springback prediction for incremental sheet forming based on FEM-PSONN technology

In the incremental sheet forming （ISF） process, springback is a very important factor that affects the quality of parts. Predicting and controlling springback accurately is essential for the design of the toolpath for ISF. A three-dimensional elasto-plastic finite element model （FEM） was developed to simulate the process and the simulated results were compared with those from the experiment. The springback angle was found to be in accordance with the experimental result, proving the FEM to be effective. A coupled artificial neural networks （ANN） and finite element method technique was developed to simulate and predict springback responses to changes in the processing parameters. A particle swarm optimization （PSO） algorithm was used to optimize the weights and thresholds of the neural network model. The neural network was trained using available FEM simulation data. The results showed that a more accurate prediction of s!oringback can be acquired using the FEM-PSONN model.

α射线露点传感器非线性修正

介绍了采用半导体探测器和温度传感器研制成的α射线露点传感器的工作原理.分析了α射线露点传感器的温度特性,表明测量范围较宽时,传感器的输出易受环境温度的影响,并且成非线性.提出一种基于神经网络高精度线性化子块网络集成插值实现露点传感温度补偿方法.利用神经网络共轭梯度算法具有逼近任意非线性函数的特点,通过训练使神经网络建立在不同环境温度下传感器输出与其实际感受的电压值之间的非线性映射关系,实现α射线露点传感器温度补偿.实验结果表明,该方法不仅能有效地消除温度的影响,而且能在神经网络的输出端得到期望的线性输出.

Neural-Network-Based Charge Density Quantum Correction of Nanoscale MOSFETs

For the treatment of the quantum effect of charge distribution in nanoscale MOSFETs,a quantum correction model using Levenberg-Marquardt back-propagation neural networks is presented that can predict the quantum density from the classical density. The training speed and accuracy of neural networks with different hidden layers and numbers of neurons are studied. We conclude that high training speed and accuracy can be obtained using neural networks with two hidden layers,but the number of neurons in the hidden layers does not have a noticeable effect, For single and double-gate nanoscale MOSFETs, our model can easily predict the quantum charge density in the silicon layer,and it agrees closely with the Schrodinger-Poisson approach.

DNA Sequence Classification Based on the Side Chain Radical Polarity of Amino Acids

The features of DNA sequence fragments were extracted from the distribution density of the condons in the individual cases of DNA sequence fragments. Based on the polarity of side chain radicals of amino acids molecules, the amino acids were classified into five categories, and the frequencies of these five categories were calculated. This kind of feature extraction based on the biological meanings not only took the content of basic groups into consideration, but also considered the marshal ing sequence of the basic groups. The hierarchical clustering analysis and BP neural network were used to classify the DNA sequence fragments. The results showed that the classification results of these two kinds of algo-rithms not only had high accuracy, but also had high consistence, indicating that this kind of feature extraction was superior over the traditional feature extraction which only took the features of basic groups into consideration.

Solving a Model of Inversion Layer Quantization Effects in Deep Submicron MOSFETs with Artificial Neural Networks

Approximate the solution of a model for inversion layer quantization effects in deep submicron MOSFETs with feed-forward artificial neural networks (ANNs) is proposed.To realize this,the solution of eigenvalue problems actually need to be considered for differential and integrodifferential operators,using ANNs.To validate the method and verify its accuracy,it is applied to the Schr o ¨dinger equation for the Morse potential problem that has an analytically known solution.Then a model is proceeded with which approximates the Schr o ¨dinger equation and the Poisson equation problem called the triangular-potential approximation.In conclusion,the presented method is simple to implement,and have several verification applications.

Improved Roberts operator for detecting surface defects of heavy rails with superior precision and efficiency

An experimental platform accompanying with the improved Roberts algorithm has been developed to achieve accurate and real-time edge detection of surface defects on heavy rails.Detection results of scratching defects show that the improved Roberts operator can attain accurate positioning to defect contour and get complete edge information.Meanwhile,a decreasing amount of interference noises as well as more precise characteristic parameters of the extracted defects can also be confirmed for the improved algorithm.Furthermore,the BP neural network adopted for defects classification with the improved Roberts operator can obtain the target training precision with 98 iterative steps and time of 2s while that of traditional Roberts operator is 118 steps and 4s.Finally,an enhanced defects identification rate of 13.33%has also been confirmed after the Roberts operator is improved.The proposed detecting platform will be positive in producing high-quality heavy rails and guaranteeing the national transportation safety.

A novel internet traffic identification approach using wavelet packet decomposition and neural network

Internet traffic classification plays an important role in network management, and many approaches have been proposed to classify different kinds of internet traffics. A novel approach was proposed to classify network applications by optimized back-propagation (BP) neural network. Particle swarm optimization (PSO) algorithm was used to optimize the BP neural network. And in order to increase the identification performance, wavelet packet decomposition (WPD) was used to extract several hidden features from the time-frequency information of network traffic. The experimental results show that the average classification accuracy of various network applications can reach 97%. Moreover, this approach optimized by BP neural network takes 50% of the training time compared with the traditional neural network.

Neural Network Based on Quantum Chemistry for Predicting Melting Point of Organic Compounds

The melting points of organic compounds were estimated using a combined method that includes a backpropagation neural network and quantitative structure property relationship （QSPR） parameters in quantum chemistry. Eleven descriptors that reflect the intermolecular forces and molecular symmetry were used as input variables. QSPR parameters were calculated using molecular modeling and PM3 semi-empirical molecular orbital theories. A total of 260 compounds were used to train the network, which was developed using MatLab. Then, the melting points of 73 other compounds were predicted and results were compared to experimental data from the literature. The study shows that the chosen artificial neural network and the quantitative structure property relationships method present an excellent alternative for the estimation of the melting point of an organic compound, with average absolute deviation of 5%.

Identification of rice seed varieties using neural network

A digital image analysis algorithm based color and morphological features was developed to identify the six varieties （ey7954, syz3, xsl 1, xy5968, xy9308, z903） rice seeds which are widely planted in Zhejiang Province. Seven color and fourteen morphological features were used for discriminant analysis, Two hundred and forty kernels used as the training data set and sixty kernels as the test data set in the neural network used to identify rice seed varieties. When the model was tested on the test data set, the identification accuracies were 90.00%, 88.00%, 95.00%, 82.00%, 74.00%, 80.00% for ey7954, syz3, xsl1, xy5968, xy9308, z903 respectively.