Determination of land surface temperature using precipitable water based Split-Window and Artificial Neural Network in Turkey

Land surface temperature (LST) calculation utilizing satellite thermal images is very difficult due to the great temporal variance of atmospheric water vapor in the atmosphere which strongly affects the thermal radiance incoming to satellite sensors. In this study, Split-Window (SW) and Radial Basis Function (RBF) methods were utilized for prediction of LST using precipitable water for Turkey. Coll 94 Split-Window algorithm was modified using regional precipitable water values estimated from upper-air Radiosond observations for the years 1990–2007 and Local Split-Window (LSW) algorithms were generated for the study area. Using local algorithms and Advanced Very High Resolution Radiometer (AVHRR) data, monthly mean daily sum LST values were calculated. In RBF method latitude, longitude, altitude, surface emissivity, sun shine duration and precipitable water values were used as input variables of the structure. Correlation coefficients between estimated and measured LST values were obtained as 99.23% (for RBF) and 94.48% (for LSW) at 00:00 UTC and 92.77% (for RBF) and 89.98% (for LSW) at 12:00 UTC. These meaningful statistical results suggest that RBF and LSW methods could be used for LST calculation.     

Intra-annual variations of the thermospheric density at 400 km altitude from 1996 to 2006

We investigate the intra-annual variations of globally averaged thermospheric density at 400 km altitude from 1996 to 2006 by using Artificial Neural Network Method (ANNM). The results indicate that thermospheric density is governed by solar activity, and the absolute error of our model is 13.67%, less than NRLMSISE-00 model. Fourier representation can catch the intra-annual variations more accurately than NRLMSISE-00 model and JB2008 model especially during 2002. We find that the Autumn maximum is slightly greater than Spring maximum during solar minimum, while the reverse is correct during solar maximum. There is a strong linear relation between solar activity and the amplitude of annual/semiannual variations, and the correlation coefficients are 0.9534 and 0.9424, respectively. Moreover, the amplitude ratio of the annual to semiannual variation is about 1.3 averaged, and changes in different years, but it has little relation with solar activity. Besides that, the amplitude of annual variation is larger than semiannual variation during 1996 and 2006 except 1998 and 2000. The relative error of NRLMSISE-00 model is 14.95%, decreasing to 12.49% after revising, and the correlation coefficients between this empirical model and its improved results and the observation are 0.8185 and 0.9210, respectively. Finally, we suggest the revised version of MSIS series of model should use the Fourier representation to express the intra-annual variations.     

基于进化神经网络的组合导航系统滤波算法研究

组合导航系统中,卡尔曼滤波算法的估计特性依赖于精确的系统模型和准确的外观测数据.任何一个条件不满足将导致滤波精度下降甚至发散.为此,本文引入进化神经网络和自适应卡尔曼滤波技术.仿真结果证明,文中所提算法能够有效克服常规卡尔曼滤波的缺点,并保持较高的精度.     

光纤陀螺温度建模及补偿技术研究

研究了干涉式光纤陀螺的温度补偿问题。首先分析了光纤陀螺仪温度漂移的主要影响因素，并应用人工神经网络与数理统计方法建立了陀螺仪的温度补偿模型。最后对某型号陀螺进行了0～50℃环境温度下的大量恒温及变温实验，通过实验数据完善所建模型并完成了对该陀螺的温度补偿。实验结果表明：在特定温度环境下，该方法能够有效地改善光纤陀螺仪零偏稳定性。     

模拟电路故障诊断新方法

在BP型ANN网络用于模拟电路故障诊断的基础上,提出了利用GA辅助设计的BP型ANN实现模拟电路故障诊断的新方法。实例验证表明,新方法的诊断精度、诊断速度以及建立诊断模型的自动化程度都有了较大的提高。     

BP网络局部极小值的修正算法

在人工神经网络中,反向传播法占据了非常重要的地位。近几十年许多研究人员对其做了深入的研究,提出了很多改进的方法,主要目标是为了加快训练速度,避免陷入局部极小值和改善其他能力。本文主要讨论在MATLAB下使用附加动量法使网络避免陷入局部极小值问题。     

基于网络流仿真的潜通路分析方法

针对线索表法在分析潜通路问题过程中约束条件过多的问题,在人工神经网络分析法的基础上应用网络流仿真进行潜通路分析.根据电路元件的电气特性以及人工神经网络的特点,建立了元件的定性仿真模型,确定电路网络的构成方式.通过网络流仿真模拟电流在电路系统中的扩散过程,预测电路中负载的响应.通过对比电路网络中负载的设计响应以及通过分析预测得到的负载响应,就可以判断出电路网络是否存在潜通路问题,并且找到发生潜通路问题的原因.该方法可以正确预测电路网络中的负载响应,克服了线索表法的缺陷,减少对分析已知条件的要求以及人为因素对分析结果的影响.     

基于人工神经网络的TCAS板卡深度故障诊断系统

提出了一种基于人工神经网络的故障定位算法,用于TCAS板卡深度故障诊断系统。     

利用人工神经网络提前1h预报电离层TEC

提出了一种利用人工神经网络提前1h预报电离层TEC的简便方法. 考虑到实际工程应用要求, 没有使用其他空间天气参数, 而是选择电离层TEC观测数据本身作为输入参数. 输入参数为当前时刻TEC、一阶差分、相对差分和时间, 输出参数为预报时刻TEC. 利用文中介绍的GPS/TEC处理方法解算厦门站2004年电离层TEC观测数据, 对预报方法进行评估, 全年平均相对误差为9.3744%, 预报结果与观测值相关性达到了0.96678. 结果表明, 利用人工神经网络方法提前1h预报电离层TEC有很好的应用前景.