基于特征提取相似日的ELM短期负荷预测研究

为解决短期电力负荷预测中预测精度差、计算时间长等问题,提出一种基于自组织特征映射网络进行特征提取相似日的极限学习机短期电力负荷预测方法。通过自组织特征映射网络找出与预测日同类型的历史数据作为训练样本;并采用预测能力强、计算时间短的ELM网络进行预测。以某市电力负荷数据进行仿真,并将上述方法与传统神经网络进行对比。仿真算例表明,基于特征提取相似日的ELM方法具有较高的预测精度,泛化性能好,且运算时间短。     

基于RDSNet的毫米波雷达人体跌倒检测方法

随着人口老龄化的到来,跌倒检测逐渐成为研究热点。针对基于毫米波雷达的人体跌倒检测应用,该文提出了一种融合卷积神经网络和长短时记忆网络的距离多普勒热图序列检测网络(RDSNet)模型。首先通过卷积神经网络对距离多普勒热图进行特征提取得到特征向量,然后将动态序列对应的特征向量序列依次输入长短时记忆网络,进而学习得到热图序列的时间相关性信息,最后通过分类器网络得到检测结果。利用毫米波雷达采集了不同对象的多种人体动作,构建了距离多普勒热图数据集。对比试验表明,所提出的RDSNet网络模型检测准确率可达到96.67%,计算时延小于50 ms,而且具有良好的泛化能力,可为跌倒检测和人体姿态识别提供新的技术思路。      

深度PCANet模型的人脸美丽预测

深度卷积神经网络（Deep Convolution Neural Network, DCNN）在人脸识别、图像分类和目标检测领域已取得较好效果,并得到广泛应用;但是,在人脸美丽预测中却存在拟合效果欠佳、网络训练难度大等问题。深度PCANet模型,将深度主元分析网络（Principal Component Analysis Network,PCANet）作为特征提取器;采用无监督预训练提取网络参数,具有网络训练时间短、图像特征提取快等特点,能有效避免DCNN存在的问题。为此,本文将深度PCANet引入人脸美丽预测,对训练集图像采用多尺度预处理,训练深度PCANet。该模型可提取人脸图像的结构性全局特征,采用特征增强方法可生成更具表征能力的特征;运用线性支持向量机（Support Vector Machine, SVM）和随机森林（Random Forest, RF）回归器进行训练和预测。基于SCUT-FBP人脸美丽数据库的实验结果表明,深度PCANet模型具有结构简单、特征提取快和无需网络调参优化等特点;选择合适的图像尺度与采用特征增强方法可提高人脸美丽评价结果,证明了所提方法的有效性和可行性。      

基于LSTM网络和特征融合的通信干扰识别

针对现有通信干扰信号识别方法识别效果不佳的问题,提出了一种基于长短时记忆网络(Long Short-Term Memory,LSTM)和特征融合的通信干扰识别方法.该方法利用LSTM网络提取干扰信号的特征,通过LSTM强大的序列特征提取能力提升干扰信号特征提取的性能;通过提取信号的时域和频域特征后进行特征融合,使用全连...     

A life-prediction method for lithium-ion batteries based on a fusion model and an attention mechanism

The current life-prediction models for lithium-ion batteries have several problems, such as the construction of complex feature structures, a high number of feature dimensions, and inaccurate prediction results. To overcome these problems, this paper proposes a deep-learning model combining an autoencoder network and a long short-term memory network. First, this model applies the characteristics of the autoencoder to reduce the dimensionality of the high-dimensional features extracted from the battery data set and realize the fusion of complex time-domain features, which overcomes the problems of redundant model information and low computational efficiency. This model then uses a long short-term memory network that is sensitive to time-series data to solve the long-path dependence problem in the prediction of battery life. Lastly, the attention mechanism is used to give greater weight to features that have a greater impact on the target value, which enhances the learning effect of the model on the long input sequence. To verify the efficacy of the proposed model, this paper uses NASA''s lithium-ion battery cycle life data set.     

基于信号能量分布拟合优度的长短时记忆网络频谱感知算法研究

传统频谱感知算法性能在低信噪比下不够理想,在高信噪比下较好,算法性能随信噪比降低逐渐变差。本文提出了基于信号能量分布拟合优度的长短时记忆网络频谱感知算法,利用授权用户信号存在时的接收信号为基础,计算接收信号的能量分布,并将通过拟合优度算法得到的距离值作为特征构造特征向量,然后将特征向量输入长短时记忆网络训练得到模型,最后将测试数据输入训练模型进行预测,从而实现频谱感知。仿真结果表明,本文提出的新算法在信噪比为-13 dB,采样点数为28时,检测概率达到96.21%,明显优于传统能量检测算法和传统拟合优度算法。      

Dual attention interactive fine-grained classification network based on data augmentation

The key to fine-grained image classification is to find discriminative regions. Most existing methods only use simple baseline networks or low-recognition attention modules to discover object differences, which will limit the model to finding discriminative regions hidden in images. This article proposes an effective method to solve this problem. The first is a novel layered training method, which uses a new training method to enhance the feature extraction ability of the baseline model. The second step focuses on key regions of the image based on improved long short-term memory (LSTM) and multi-head attention. In the third step, based on the feature map obtained by the dual attention network, spatial mapping is performed by a multi-layer perceptron (MLP). Then the element-by-element mutual multiplication calculation of the channel is performed to obtain a feature map with finer granularity. Finally, the CUB-200-2011, FGVC Aircraft, Stanford Cars, and MedMNIST v2 datasets achieved good performance.     

基于小波变换和神经网络的短期风电功率预测方法

为了降低电网调度的操作复杂性,同时提升风电功率的预测精度,本文主要分析影响风电功率预测方法的因素,包括风向、风速以及环境温度等方面,同时经过对比研究,小波变换和神经网络的短期风电功率预测方法能够提升预测精度.     

Research on EEG emotion recognition based on CNN+BiLSTM+self-attention model

To address the problems of insufficient dimensionality of electroencephalogram (EEG) feature extraction, the tendency to ignore the importance of different sequential data segments, and the poor generalization ability of the model in EEG based emotion recognition, the model of convolutional neural network and bi-directional long short-term memory and self-attention (CNN+BiLSTM+self-attention) is proposed. This model uses convolutional neural network (CNN) to extract more distinctive features from both spatial and temporal dimensions. The bi-directional long short-term memory (BiLSTM) is used to further preserve the long-term dependencies between the temporal phases of sequential data. The self-attention mechanism can change the weights of different channels to extract and highlight important information and address the often-ignored importance of different channels and samples when extracting EEG features. The subject-dependent experiment and subject-independent experiment are performed on the database for emotion analysis using physiological signals (DEAP) and collected datasets to verify the recognition performance. The experimental results show that the model proposed in this paper has excellent recognition performance and generalization ability.     

基于卷积神经网络的网络隐写分析方法研究

针对现有网络隐写分析算法特征提取难度大、算法适用范围单一的问题,文章提出了一种基于卷积神经网络的网络隐写分析方法。对网络数据流进行预处理,将所有数据包处理成大小相同的矩阵,最大限度地保留数据特征完整性;使用异构卷积进行特征提取,减少模型计算量及参数数量,加快模型收敛速度;取消池化层,提高模型训练效率。与传统网络隐写分析方法相比,模型能够自动提取数据特征,识别多种网络隐写算法。     

基于神经网络与马尔可夫组合模型的视频流行度预测算法

为了提升用户体验,降低运营商的成本,将播放最多的视频内容提前放入用户侧缓存是业界的通用做法,如何有效预测视频播放热度已经成为业界热点问题。针对传统预测算法非线性映射能力差、预测精度低及自适应性弱等缺点,提出基于神经网络与马尔可夫组合模型的视频流行度预测算法（Mar-BiLSTM）,该算法通过构建双向长短期记忆（bi-directional long short-term memory,BiLSTM）网络模型可以保留时间序列两个方向的信息依赖;同时在避免引入外部变量导致模型复杂度增加的情况下,利用马尔可夫性质进一步提高了模型的预测精度。实验结果表明,与传统的时间序列和经典的神经网络算法相比,所提算法提升了视频流行度预测的准确性、时效性,并降低了计算量。     

Exo-LSTM: traffic flow prediction based on multifractal wavelet theory

In order to predict traffic flow more accurately and improve network performance, based on the multifractal wavelet theory, a new traffic prediction model named exo-LSTM is proposed. Exo represents exogenous sequence used to provide a detailed sequence for the model, LSTM represents long short-term memory used to predict unstable traffic flow. Applying multifractal traffic flow to the exo-LSTM model and other existing models, the experiment result proves that exo-LSTM prediction model achieves better prediction accuracy.     

一种超宽带与惯导融合的LSTM室内定位算法

针对复杂环境下的室内高精度定位需求,提出了一种超宽带和惯导融合定位方案。结合位置估计过程可被划分为时间序列预测问题的特点,提出了一种基于长短时记忆(Long Short Term Memory,LSTM)网络的联合定位算法,并对其总体架构设计、数据预处理方法、网络结构设计、模型训练方法进行了研究。在此基础上,通过仿真和实测实验对联合定位算法进行验证,实验结果表明,该LSTM神经网络联合定位算法的定位精度优于传统TOA(Time of Arrival)、UKF(Unscented Kalman Filter)联合定位算法,适用复杂室内定位。     

基于VMD和DAIPSO-GPR解决容量再生现象的锂离子电池寿命预测研究

锂离子电池应用时表现出的时变、动态、非线性等特征,以及容量再生现象,导致传统模型对锂离子电池剩余使用寿命(RUL)预测的准确性低,该文将变分模态分解(VMD)和高斯过程回归(GPR)以及动态自适应免疫粒子群(DAIPSO)结合,建立RUL预测模型。首先利用等压降放电时间分析法,提取健康因子,利用VMD对其进行分解处理,挖掘数据内在信息,降低数据复杂度,并针对不同分量,利用不同协方差函数建立GPR预测模型,有效捕获了数据的长期下降趋势和短期再生波动。利用DAIPSO算法优化GPR模型,实现核函数超参数的优化,建立了更准确的退化关系模型,最终实现剩余使用寿命的准确预测,以及不确定性表征。最后利用NASA电池数据进行验证,离线预测结果表明所提方法具有较高预测精度和泛化适应能力。     

Remaining useful life prediction of lithium-ion batteries using a fusion method based on Wasserstein GAN

Lithium-ion batteries are the main power supply equipment in many fields due to their advantages of no memory, high energy density, long cycle life and no pollution to the environment. Accurate prediction for the remaining useful life (RUL) of lithium-ion batteries can avoid serious economic and safety problems such as spontaneous combustion. At present, most of the RUL prediction studies ignore the lithium-ion battery capacity recovery phenomenon caused by the rest time between the charge and discharge cycles. In this paper, a fusion method based on Wasserstein generative adversarial network (GAN) is proposed. This method achieves a more reliable and accurate RUL prediction of lithium-ion batteries by combining the artificial neural network (ANN) model which takes the rest time between battery charging cycles into account and the empirical degradation models which provide the correct degradation trend. The weight of each model is calculated by the discriminator in the Wasserstein GAN model. Four data sets of lithium-ion battery provided by the National Aeronautics and Space Administration (NASA) Ames Research Center are used to prove the feasibility and accuracy of the proposed method.     

基于深度学习的海上目标一维序列信号目标检测方法

当前海面目标检测方法多基于统计理论,检测性能受背景统计特性假设的影响,本文从信号预测和特征分类两个角度,分别采用长短时记忆网络(LSTM)和卷积神经网络(CNN)对信号时间序列幅度信息进行处理,用于海上目标一维序列雷达信号检测,该方法不需事先假设背景统计特性,泛化能力更强。基于LSTM序列预测的目标检测方法通过用海杂波信号幅度时间序列对网络进行训练,再用训练后的网络对后续序列进行预测,并与后续实测信号进行比较,实现目标检测。基于CNN序列分类的目标检测方法中采用截取的海杂波信号和目标信号幅度序列作为数据集样本,对一维卷积核CNN进行训练,使其具有识别目标杂波信号特征能力,从而实现目标检测。最后,采用IPIX和CSIR实测海杂波数据对两种方法进行验证,结果表明两种方法均可实现一维序列信号中海面目标的检测,但LSTM预测方法对于长序列检测的实时性有待于进一步提高;CNN分类方法可实现实时检测,但仅利用信号幅度信息,检测性能仍需进一步提升。      

基于轻量级注意机制的人脸检测算法

提出一个新的基于轻量级注意力机制的网络框架。在YOLOv3主干网络的基础上,使用深度卷积和点卷积代替标准卷积设计特征提取网络,加快模型的训练,提高检测的速度,然后引入注意力机制模块进行模型速度和精度的权衡,最后通过增加多尺度提取更多网络层的特征信息,同时使用K-means++聚类算法进一步优化网络参数。实验结果表明,该方法可以显著提高人脸检测模型的性能,在Wider Face数据集上可以达到94.08%的准确率和83.97%的召回率,且平均检测时间只需0.022 s,相比原始YOLOv3算法提高了4.45倍。     

基于贝叶斯深度学习的一维雷达有源干扰信号识别方法

雷达干扰信号准确识别是雷达抗干扰的前提,对于雷达生存至关重要。针对传统雷达干扰信号识别方法需要繁琐的分析计算提取特征,通用性差,泛化能力弱,难以适应复杂的雷达工作环境问题。本文考虑无需人工提取特征信息且具有较好的分类识别效果的深度学习网络。考虑到传统的深度学习网络由于使用点估计方式,不能够很好的衡量预测结果中的不确定性,本文提出了一种基于贝叶斯深度学习的干扰识别方法。首先,通过概率建模代替网络参数模型的点估计,解决了不确定性随机数据引起的网络过拟合问题。其次,考虑有效利用雷达回波信号的时序特性设计了LSTM层,同时解决训练过程中的梯度消失问题。基于线性调频雷达有源干扰实测数据完成了网络训练与测试,实验结果表明,引入贝叶斯方法可以在加快网络收敛速度的同时有效提高识别准确率。     

An intelligent maximum power extraction algorithm for inverter-based variable speed wind turbine systems

This paper focuses on the development of maximum wind power extraction algorithms for inverter-based variable speed wind power generation systems. A review of existing maximum wind power extraction algorithms is presented in this paper, based on which an intelligent maximum power extraction algorithm is developed by the authors to improve the system performance and to facilitate the control implementation. As an integral part of the max-power extraction algorithm, advanced hill-climb searching method has been developed to take into account the wind turbine inertia. The intelligent memory method with an on-line training process is described in this paper. The developed maximum wind power extraction algorithm has the capability of providing initial power demand based on error driven control, searching for the maximum wind turbine power at variable wind speeds, constructing an intelligent memory, and applying the intelligent memory data to control the inverter for maximum wind power extraction, without the need for either knowledge of wind turbine characteristics or the measurements of mechanical quantities such as wind speed and turbine rotor speed. System simulation results and test results have confirmed the functionality and performance of this method.     

利用通道剪枝技术的实时实例分割方法

随着实例分割技术在各种场景中的应用越来越广泛,运行速度和硬件资源占用是该技术在应用中需要考虑的2个重要因素。最近提出的基于图像原型掩码系数的实例分割网络(YOLACT)在运行速度方面做得很好,但是需要设置较大的特征提取网络才能保证分割精确度,这就导致了模型占用的硬件资源较多,同时运行速度也受到了限制。在YOLACT的基础上,提出一种新的模型,对实例分割的特征提取网络进行了优化,先使用基于批量归一化层放缩因子的通道剪枝方法对YOLACT网络进行压缩,然后对压缩后的卷积层和批量归一化层进行融合,最后,在COCO val2017上对本文提出的方法进行了评估。实验结果表明,相比原始的YOLACT网络,该方法的模型文件大小可以减少56.9%,运行速度提升28.6%,运行时显存占用也降低了13.6%,有效地减少了硬件资源占用,并且提升了运行速度。