IUKF neural network modeling for FOG temperature drift

A novel neural network based on iterated unscented Kalman filter （IUKF） algorithm is established to model and com- pensate for the fiber optic gyro （FOG） bias drift caused by temperature. In the network, FOG temperature and its gradient are set as input and the FOG bias drift is set as the expected output. A 2-5-1 network trained with IUKF algorithm is established. The IUKF algorithm is developed on the basis of the unscented Kalman filter （UKF）. The weight and bias vectors of the hidden layer are set as the state of the UKF and its process and measurement equations are deduced according to the network architecture. To solve the unavoidable estimation deviation of the mean and covariance of the states in the UKF algorithm, iterative computation is introduced into the UKF after the measurement update. While the measure- ment noise R is extended into the state vectors before iteration in order to meet the statistic orthogonality of estimate and mea- surement noise. The IUKF algorithm can provide the optimized estimation for the neural network because of its state expansion and iteration. Temperature rise （-20-20℃） and drop （70-20℃） tests for FOG are carried out in an attemperator. The temperature drift model is built with neural network, and it is trained respectively with BP, UKF and IUKF algorithms. The results prove that the proposed model has higher precision compared with the back- propagation （BP） and UKF network models.

Optimization of Laser Ablation Technology for PDPhSM Matrix Nanocomposite Thin Film by Artificial Neural Networks-particle Swarm Algorithm

A new thermal ring-opening polymerization technique for 1, 1, 3, 3-tetra-ph enyl-1, 3-disilacyclobutane （TPDC） based on the use of metal nanoparticles produced by pulsed laser ablation was investigated. This method facilitates the synthesis of polydiphenysilylenemethyle （PDPhSM） thin film, which is difficult to make by conventional methods because of its insolubility and high melting point. TPDC was first evaporated on silicon substrates and then exposed to metal nanoparticles deposition by pulsed laser ablation prior to heat treatment.The TPDC films with metal nanoparticles were heated in an electric furnace in air atmosphere to induce ring-opening polymerization of TPDC. The film thicknesses before and after polymerization were measured by a stylus profilometer. Since the polymerization process competes with re-evaporation of TPDC during the heating, the thickness ratio of the polymer to the monomer was defined as the polymerization efficiency, which depends greatly on the technology conditions. Therefore, a well trained radial base function neural network model was constructed to approach the complex nonlinear relationship. Moreover, a particle swarm algorithm was firstly introduced to search for an optimum technology directly from RBF neural network model. This ensures that the fabrication of thin film with appropriate properties using pulsed laser ablation requires no in-depth understanding of the entire behavior of the technology conditions.

EFFECT OF COLD WORKING ON THE AGING PROPERTIES OF Cu-Cr-Zr-Mg ALLOY BY ARTIFICIAL NEURAL NETWORK

A developmental research has been carried out to deal with the high performance of Cu-Cr-Zr-Mg lead frame alloy by artificial neural network (ANN). Using the cold working to assist in the aging hardening can improve the the hardness and electrical conductivity properties of Cu-Cr-Zr-Mg lead frame alloy. This paper studies the effect of different extent of cold working on the aging properties by a supervised ANN to model the non-linear relationship between processing parameters and the properties. The back-propagation (BP) training algorithm is improved by Levenberg-Marquardt algorithm. A basic repository on the domain knowledge of cold worked aging processes is established via sufficient data mining by the network. The predicted values of the ANN coincide well with the tested data. So an important foundation has been laid for prediction and optimum controlling the rolling and aging properties of Cu-Cr-Zr-Mg alloy.

Interaction Energy Prediction of Organic Molecules using Deep Tensor Neural Network

The interaction energy of two molecules system plays a critical role in analyzing the interacting effect in molecular dynamic simulation.Since the limitation of quantum mechanics calculating resources,the interaction energy based on quantum mechanics can not be merged into molecular dynamic simulation for a long time scale.A deep learning framework,deep tensor neural network,is applied to predict the interaction energy of three organic related systems within the quantum mechanics level of accuracy.The geometric structure and atomic types of molecular conformation,as the data descriptors,are applied as the network inputs to predict the interaction energy in the system.The neural network is trained with the hierarchically generated conformations data set.The complex tensor hidden layers are simplified and trained in the optimization process.The predicted results of different molecular sys tems indica te that deep t ensor neural net work is capable to predic t the interaction energy with 1 kcal/mol of the mean absolute error in a relatively short time.The prediction highly improves the efficiency of interaction energy calculation.The whole proposed framework provides new insights to introducing deep learning technology into the interaction energy calculation.

Sound Quality Prediction of Vehicle Interior Noise under Multiple Working Conditions Using Back-Propagation Neural Network Model

This paper presents a back-propagation neural network model for sound quality prediction (BPNN-SQP) of multiple working conditions’ vehicle interior noise. According to the standards and regulations, four kinds of vehicle interior noises under operating conditions, including idle, constant speed, accelerating and braking, are acquired. The objective psychoacoustic parameters and subjective annoyance results are respectively used as the input and output of the BPNN-SQP model. With correlation analysis and significance test, some psychoacoustic parameters, such as loudness, A-weighted sound pressure level, roughness, articulation index and sharpness, are selected for modeling. The annoyance values of unknown noise samples estimated by the BPNN-SQP model are highly correlated with the subjective annoyances. Conclusion can be drawn that the proposed BPNN-SQP model has good generalization ability and can be applied in sound quality prediction of vehicle interior noise under multiple working conditions.

Residential Community Open-Up Strategy Based on Prim’s Algorithm and Neural Network Algorithm

“Open community” has aroused widespread concern and research. This paper focuses on the system analysis research of the problem that based on statistics including the regression equation fitting function and mathematical theory, combined with the actual effect of camera measurement method, Prim’s algorithm and neural network to “Open community” and the applicable conditions. Research results show that with the increasing number of roads within the district, the benefit time gradually increased, but each type of district capacity is different.

基于FMCW雷达的人体生命体征信号预测算法

将FMCW雷达检测到的人体生命体征信号,用于预测未来一段时间内人体生命体征信号是否异常,具有明显的应用价值。该方向当前研究主要针对如何进一步降低重构误差、提升生命体征信号的预测精度。为此,本文提出一种自适应变分模态分解-长短期记忆神经网络的生命体征信号预测方法。针对静止状态下的人体,通过雷达采集到的生命体征信号,采用粒子群算法优化变分模态分解VMD的模态分量个数K和惩罚系数α的值,实现自适应选取后用于VMD分解,再将分解后的模态分量进行叠加重构。采用粒子群算法优化长短期记忆网络模型中的网络层数、学习率、正则化系数等3个参数,自适应选取合适的参数组合,将重构后的信号通过优化后的LSTM网络进行预测。实验结果显示本文所提预测方法在10位志愿者的预测结果与原始数据的均方根误差平均值为0.017 188 9,平均绝对误差的平均值为0.007 158,相较于当前其他研究,预测精度上有明显提升。

An Intelligent System for Recognition of the Work Wave of an EC Engine Based on a Neural Network

This paper introduces the principle for recognition of engine work wave signal with neural network. A diagnosis method for recognizing engine trouble by its work wave is proposed. The designing process is illustrated by diagnosing the voltage trouble of the fuel injector of an electronic control (EC) engine.

基于边缘辅助和多尺度Transformer的无参考屏幕内容图像质量评估

与从现实场景中拍摄的自然图像不同,屏幕内容图像是一种合成图像,通常由计算机生成的文本、图形和动画等各种多媒体形式组合而成.现有评估方法通常未能充分考虑图像边缘结构信息和全局上下文信息对屏幕内容图像质量感知的影响.为解决上述问题,本文提出一种基于边缘辅助和多尺度Transformer的无参考屏幕内容图像质量评估模型.首先,使用高斯拉普拉斯算子构造由失真屏幕内容图像高频信息组成的边缘结构图,然后通过卷积神经网络(Convolutional Neural Network,CNN)对输入的失真屏幕内容图像和相应的边缘结构图进行多尺度的特征提取与融合,以图像的边缘结构信息为模型训练提供额外的信息增益.此外,本文进一步构建了基于Transformer的多尺度特征编码模块,从而在CNN获得的局部特征基础上更好地建模不同尺度图像和边缘特征的全局上下文信息.实验结果表明,本文提出的方法在指标上优于其他现有的无参考和全参考屏幕内容图像质量评估方法,能够取得更高的主客观视觉感知一致性.

基于混合神经网络的多维视觉传感信号模式分类

传感器采集的数字信号分类精度差,导致关键信息的丢失。为了提高传感数据的可靠性和有效性,提出基于混合神经网络的多维视觉传感信号模式分类方法。结合卷积神经网络(CNN)、循环神经网络(RNN)构建混合神经网络,以更有效地表示多维视觉数据中的特征;其中,卷积神经网络负责对多维的空间信号进行去噪处理并提取特征;循环神经网络负责对时域和频域信号进行特征提取;混合神经网络通过联合训练CNN和RNN各自的参数,以调整其权重,并且结合两者从不同层级提取的特征来实现多维视觉传感信号模式的分类。仿真结果表明,使用所提方法进行分类时,信号光滑度保持在0.9以上,传感信号分类结果与实际结果拟合度较高,有效实现多维视觉传感信号模式分类。

联合图像层级特征的压缩感知迭代重构

基于卷积神经网络(Convolutional Neural Networks,CNN)的图像压缩感知重构算法难以捕捉高分辨率图像的长距离依赖关系,采用Transformer虽能解决该问题,但网络参数量和图像重构时间成倍增长。基于此,本文提出了一种联合图像层级特征的压缩感知迭代重构网络(Combining Image Hierarchical-Feature Network,CHFNet),在提高图像重构质量的同时减少重构时间。CHFNet由采样和重构两个子网络组成,采样子网络通过可学习的采样矩阵为重构过程提供更有效的测量值。在重构子网络中,设计了一种使用梯度下降操作和特征优化操作的迭代策略,同时提出一种轻量级CNN-Transformer混合架构,能够建模并优化高细粒度的图像层级特征,在增强网络感知能力的同时降低计算复杂度。此外,CHFNet通过联合优化学习采样重构,实现了完整的端到端训练。实验结果表明,所提算法在多个公共基准数据集上取得了良好的重构效果。在Urban100数据集上,相较于现有最优算法CSformer,平均PSNR,SSIM分别提升0.63 dB和0.0076;在0.10采样率下,相较CSformer在Set11,BSD68和Urban100数据集上的平均重构时间分别减少了2.7447 s,3.5510 s和4.7750 s。

基于卷积神经网络的多工况多传感滚动轴承实时监控方法

针对工业环境中广泛在多工况下多滚动轴承实时状态监测的需求和部署环境受限的挑战,提出一种基于卷积神经网络(Convolutional Neural Network,CNN)的面向多传感器滚动轴承运行状态监控方法。该方法将两个不同工况下的一维时间序列数据集以均方根(Root Mean Square,RMS)指标标注,并通过将一维时间序列多传感器数据重构为二维空间张量的形式输入卷积神经网络训练。最后利用层融合和16比特量化优化,将网络部署到FPGA上,用以解决CNN的计算开销。实验结果表明,在结合了两种不同工况的数据集下,网络测试推理准确度依然高达99.24%,比多层感知机实现高10.48%,比多层感知机结合支持向量机的实现高2.91%,该算法对于新加入的数据集也有较强的鲁棒性,经过重训练,新加入的数据集准确率可以达到99.17%。基于FPGA部署优化的网络的峰值能效为76.217GPOS/W,为CPU实现的33.09倍,GPU实现的5.39倍。其中,16比特精度部署的网络测试精度相较32比特精度实现仅降低0.001%。

融合信息瓶颈与图卷积的跨域推荐算法

基于迁移学习的跨域推荐可以有效地学习连接源域和目标域的映射函数,但其性能仍然受到表征质量不高和负迁移问题的影响,不能有效地为冷启动用户进行推荐,为此提出了一种融合信息瓶颈与图卷积网络的跨域推荐模型(IBGC)。利用图卷积神经网络聚合有关联的用户-用户和项目-项目信息;利用注意力机制学习用户和项目偏好,以提高节点特征表示质量;考虑到两个领域的信息交互,将重叠用户进行嵌入表示的同时限制特定信息的编码,利用信息瓶颈理论设计了三种正则化器,以捕获域内和跨域用户-项目的相关性,并将不同领域的重叠用户表征对齐以解决负迁移问题。在Amazon数据集中的四对公开数据集上进行实验,实验结果表明该模型在MRR、HR@K和NDCG@K三个推荐性能指标上的表现均优于基线模型,在四对数据集上与最优对比基线模型相比,MRR平均提升34.36%,HR@10平均提升34.94%,NDCG@10平均提升36.83%,证明了IBGC模型的有效性。

基于对抗生成网络的手写数字生成方法研究

提出了一种基于改进的对抗生成网络(GANs)的手写数字生成方法。通过引入GANs的生成器和判别器,设计了一种创新的网络架构,并采用了一系列训练策略来解决GANs训练中的不稳定性和模式崩溃问题。通过使用MNIST数据集进行实验,结果显示,新模型在准确率上达到了98%以上。与传统神经网络方法相比,基于改进的对抗生成网络的手写数字识别系统在性能上有显著提升。

基于多神经网络的可展开网状天线型面调整方法

为在提高可展开网状天线型面精度的同时减少型面调整的工作量,提出了一种基于多神经网络的型面调整方法。通过分析新型张拉网状天线型面与调整索相关性与耦合机制,首次提出了型面调整策略。以10m口径的新型张拉网状天线为例进行了数值仿真研究,调整后型面的均方根值从5.4×10^(-3)m降低到1.1×10^(-3)m,从而验证了方法的有效性。

不同神经类型大学生工作记忆的特点及神经机制

为探讨不同神经类型大学生工作记忆加工的特点,用80-8神经类型量表筛选灵活型、安静型、兴奋型和抑制型四组典型的神经类型大学生并让其完成ERP版的N-back任务。结果发现灵活型大学生的正确率显著高于兴奋型和抑制型;在任务所诱发的P3波幅上,灵活型被试大于抑制型,在P3的潜伏期上不同类型之间无差异。说明灵活型个体与抑制型个体之间的差异不是来源于工作记忆加工的匹配子任务阶段,而是由于工作记忆的刷新子任务上,灵活型个体比抑制型个体加工能力更强。

基于深度神经网络的光伏发电时间序列多元预测

利用不同时间序列间的相关性和依赖性基于深度神经网络(DNNs)提出了两种不同的多元长短期记忆网络(LSTM)光伏输出功率预测方法,充分考虑了空气温度、风速等影响因素之间的相关性特征。以光伏发电站运行数据为例,通过对光伏发电预测模型进行训练和测试,并与单变量LSTM和Stacked-LSTM模型的结果进行比较,研究结果表明,所提的Conv-LSTM可以在减少30.76%训练时间的基础上提升0.71%~1.33%的准确度,Conv-LSTM和Multi-LSTM分别以高达93.12%和96.12%的准确度跟踪实际光伏发电。

基于贝叶斯神经网络的相位梯度计算方法

应变重构是相衬光学相干层析力学性能表征中的关键步骤,其需要准确计算出差分包裹相位的梯度分布。为了能够解决强噪声干扰下的相位梯度重构信噪比低的难题,提出了一种基于贝叶斯神经网络的相位梯度计算方法。首先,通过计算机模拟不同散斑噪声等级下的包裹相位图,并生成相应的理想相位梯度,以构建网络的训练集。其次,基于网络训练集采用贝叶斯推断理论学习强噪声环境下的包裹相位与相位梯度的“端到端”映射关系。最后,将相衬光学相干层析测量的差分包裹相位结果送入贝叶斯神经网络进行处理,实现高信噪比相位梯度预测。此外,通过借助贝叶斯神经网络的统计特性,以模型不确定度来定量评估相位梯度预测结果的可靠性。通过数值实验和三点弯曲力学加载实验对比分析了本文方法和主流矢量方法的性能。实验结果表明:在噪声较小的条件下,本文方法重构的相位梯度信噪比可提升8%;在噪声较强条件下,本文方法能成功恢复因相位条纹难以分辨而无法计算的相位梯度。此外,模型不确定度能够定量分析网络的相位梯度预测误差。可以预见,在样品形变复杂且先验信息未知的条件下,本工作为相衬光学相干层析提供了一种有效的应变重构方法,从而能实现高质量和高可靠的内部力学性能表征。

基于超球面对偶学习的双通道图异常检测方法

图异常检测作为一项重要的数据挖掘任务,专注于识别与大多数节点显著偏离的异常节点.随着无监督图神经网络技术的进步,现已开发出了基于密度估计、对抗生成网络等多种高效识别图数据中潜在异常的方法 .然而,这些方法更注重无监督图异常检测生成高质量的表征,而往往忽略了图异常的特性.因此,本文提出了一个双通道异构图异常检测模型(Dual-channel Heterogeneous Graph Anomaly Detection,HD-GAD).其模型基础架构包括双通道的图神经网络:全局子结构感知的图神经网络和局部子结构感知的图神经网络,用于图异常检测捕获全局和局部子结构属性.同时,基于对偶推断引入了多超球体学习目标(Multi-Hypersphere Learning,MHL),从宏观和介观超球体角度,分别测量在整个图/社区结构中偏离的异常节点. HD-GAD模型利用相似度函数EmbSim优化训练目标,以缓解多超球面学习中的模型坍问题.最后,在五种不同的数据集上进行了全面的实验.其AUC(Area Under Curve)值在大多数情况下均超过了0.9,达到了行业领先水平,进一步证明了HD-GAD模型在图异常检测任务上的高效性与性能优势.

S-CO_(2)介质止推箔片气体动压轴承特性研究

针对超临界二氧化碳(SupercriticalCarbon Dioxide,S-CO_(2))润滑波箔型止推箔片气体动压轴承,通过反向传播(Back Propagation,BP)神经网络算法提出S-CO_(2)的物性模型,并考虑轴承工作时的非理想气体效应,提出考虑湍流效应气体润滑模型、箔片结构力学模型和气膜平均温升计算方法,对止推箔片气体动压轴承的静动态特性进行研究,并分析不同结构参数对箔片气体轴承静动态特性的影响规律.结果表明,本文提出的物性模型准确度高,相关系数高达99.997%.以S-CO_(2)为润滑介质的止推箔片气体动压轴承具有更高的承载力,且在适当范围内减小最小初始气膜厚度或增加膜厚比可以提高轴承的承载力.以S-CO_(2)为介质的止推箔片气体动压轴承的动态刚度系数和动态阻尼系数均远高于常温常压空气介质下的止推箔片气体动压轴承.随着最小初始气膜厚度减小,轴承的动态刚度系数和动态阻尼系数均迅速增加.